{ "$schema": "https://modeldeprecations.dev/api/v1/schema.json", "generatedAt": "2026-08-03T12:47:14.256Z", "asOf": "2026-08-03", "count": 385, "models": [ { "provider": "amazon", "model": "nova-canvas", "name": "Amazon Nova Canvas", "aliases": [], "description": "Amazon's image generation model, capped at a 1024-character prompt and about 4.2 megapixels of output. Legacy since 30 March 2026, ending 30 September. Bedrock's lifecycle table lists no successor for it, so this is a capability leaving the platform rather than a version bump. The Bedrock invoke id is amazon.nova-canvas-v1:0.", "deprecated_on": "2026-03-30", "shutdown_on": "2026-09-30", "status": "deprecated", "replacements": [ { "provider": "google", "model": "imagen-4.0-generate-001", "recommended": false, "note": "Amazon names no successor and Nova 2 ships no image generation model, so this is a capability leaving Bedrock rather than a version bump. Imagen 4 is the nearest live equivalent in this catalogue, and it is our suggestion rather than Amazon's.", "external": false } ], "sources": [ { "url": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html", "title": "Amazon Bedrock model lifecycle", "accessed": "2026-08-03" }, { "url": "https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html", "title": "What is Amazon Nova", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "amazon", "model": "nova-lite", "name": "Amazon Nova Lite", "aliases": [], "description": "The cheap multimodal Nova, taking the same text, image and video inputs as Pro at a fraction of the price and the same 300k window. It is a distillation student of both Premier and Pro, which means the model AWS taught it with is now Legacy while the student is not. The Bedrock invoke id is amazon.nova-lite-v1:0.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html", "title": "What is Amazon Nova", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "amazon", "model": "nova-micro", "name": "Amazon Nova Micro", "aliases": [], "description": "The text-only Nova, built for the lowest latency Amazon can serve at 128k of context. It is the only understanding-tier Nova that takes no images or video, and the only one still Active in every region it launched in. The Bedrock invoke id is amazon.nova-micro-v1:0.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html", "title": "What is Amazon Nova", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "amazon", "model": "nova-premier", "name": "Amazon Nova Premier", "aliases": [], "description": "The most capable model Amazon has trained itself, a million-token multimodal system sold as much for distilling smaller custom models as for serving traffic. It went Legacy on 13 March 2026 and ends on 14 September. Anyone who used it as a distillation teacher loses the teacher, not just the endpoint. The Bedrock invoke id is amazon.nova-premier-v1:0.", "deprecated_on": "2026-03-13", "shutdown_on": "2026-09-14", "status": "deprecated", "replacements": [ { "provider": "amazon", "model": "nova-pro", "recommended": false, "note": "Bedrock's lifecycle table names no successor. Nova Pro is the most capable Active first-generation Nova; AWS has since opened a separate Nova 2 user guide, but no Nova 2 model id appears in the v1 documentation cited here.", "external": false } ], "sources": [ { "url": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html", "title": "Amazon Bedrock model lifecycle", "accessed": "2026-08-03" }, { "url": "https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html", "title": "What is Amazon Nova", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "amazon", "model": "nova-pro", "name": "Amazon Nova Pro", "aliases": [], "description": "The multimodal workhorse of the first Nova generation, 300k of context across text, image and video. With Premier now Legacy, this is the most capable Nova still Active, and the only understanding-tier model AWS lists as a distillation teacher that is not itself scheduled for end of life. The Bedrock invoke id is amazon.nova-pro-v1:0.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html", "title": "What is Amazon Nova", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "amazon", "model": "nova-reel-1-0", "name": "Amazon Nova Reel v1.0", "aliases": [], "description": "The first release of Amazon's video generation model, producing 720p at 24 frames a second from a 512-character prompt. Both Reel versions were moved to Legacy on the same day and share a 30 September 2026 end of life, so there is no newer Reel to migrate to. The Bedrock invoke id is amazon.nova-reel-v1:0.", "deprecated_on": "2026-03-30", "shutdown_on": "2026-09-30", "status": "deprecated", "replacements": [ { "provider": "minimax", "model": "MiniMax-H3", "recommended": false, "note": "Amazon names no successor and Nova 2 ships no video generation model. MiniMax H3 is the nearest live video model in this catalogue; it is our suggestion, not Amazon's.", "external": false } ], "sources": [ { "url": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html", "title": "Amazon Bedrock model lifecycle", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "amazon", "model": "nova-reel-1-1", "name": "Amazon Nova Reel v1.1", "aliases": [], "description": "The second Reel, and the id AWS's Nova documentation still lists as the current video model even though Bedrock's lifecycle table has it Legacy with a 30 September 2026 end of life. When a provider's own two pages disagree, the lifecycle table is the one with the date on it. The Bedrock invoke id is amazon.nova-reel-v1:1.", "deprecated_on": "2026-03-30", "shutdown_on": "2026-09-30", "status": "deprecated", "replacements": [ { "provider": "minimax", "model": "MiniMax-H3", "recommended": false, "note": "Amazon names no successor and Nova 2 ships no video generation model. MiniMax H3 is the nearest live video model in this catalogue; it is our suggestion, not Amazon's.", "external": false } ], "sources": [ { "url": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html", "title": "Amazon Bedrock model lifecycle", "accessed": "2026-08-03" }, { "url": "https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html", "title": "What is Amazon Nova", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "amazon", "model": "nova-sonic", "name": "Amazon Nova Sonic", "aliases": [], "description": "Amazon's speech-to-speech model, taking audio in and returning audio plus a transcript over a bidirectional stream, in five languages. Legacy since 13 March 2026 with an end of life on 14 September, and Bedrock's table names no replacement - there is currently no other Amazon-trained speech model on Bedrock. The Bedrock invoke id is amazon.nova-sonic-v1:0.", "deprecated_on": "2026-03-13", "shutdown_on": "2026-09-14", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-realtime-2.1", "recommended": false, "note": "Bedrock's lifecycle table names no successor and Nova 2 Sonic is documented by name only, with no Bedrock model id published in the pages cited here. OpenAI's realtime model is the nearest live speech-to-speech equivalent this catalogue can point at; it is our suggestion, not Amazon's.", "external": false } ], "sources": [ { "url": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html", "title": "Amazon Bedrock model lifecycle", "accessed": "2026-08-03" }, { "url": "https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html", "title": "What is Amazon Nova", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "anthropic", "model": "claude-1.0", "name": "Claude 1.0", "aliases": [], "description": "The first Claude released to developers, in March 2023, and the model that introduced Constitutional AI to a public API. It used the text-completion endpoint with literal \"Human:\" and \"Assistant:\" turn markers rather than a structured message list — a shape that reads as archaeology now.", "deprecated_on": "2024-09-04", "shutdown_on": "2024-11-06", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Two generations and a whole API separate these. Claude 1 predates the Messages API, so a migration means moving off /v1/complete and its \"\\n\\nHuman:\"/\"\\n\\nAssistant:\" prompt format entirely.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-1.1", "name": "Claude 1.1", "aliases": [], "description": "A mid-2023 refresh of the original Claude, largely a reliability and instruction-following pass rather than a capability jump. Anthropic shipped 1.0 through 1.3 within months of each other and then retired all four on a single day.", "deprecated_on": "2024-09-04", "shutdown_on": "2024-11-06", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Two generations and a whole API separate these. Claude 1 predates the Messages API, so a migration means moving off /v1/complete and its \"\\n\\nHuman:\"/\"\\n\\nAssistant:\" prompt format entirely.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-1.2", "name": "Claude 1.2", "aliases": [], "description": "The third Claude 1 point release, and the one many early integrations pinned because it was the current default when the API opened more widely. Retired November 2024 with the rest of the line.", "deprecated_on": "2024-09-04", "shutdown_on": "2024-11-06", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Two generations and a whole API separate these. Claude 1 predates the Messages API, so a migration means moving off /v1/complete and its \"\\n\\nHuman:\"/\"\\n\\nAssistant:\" prompt format entirely.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-1.3", "name": "Claude 1.3", "aliases": [], "description": "The last and most capable Claude 1, and the version most benchmark papers from 2023 actually measured. It was superseded by Claude 2 within a few months but stayed callable for over a year afterwards.", "deprecated_on": "2024-09-04", "shutdown_on": "2024-11-06", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Two generations and a whole API separate these. Claude 1 predates the Messages API, so a migration means moving off /v1/complete and its \"\\n\\nHuman:\"/\"\\n\\nAssistant:\" prompt format entirely.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-2.0", "name": "Claude 2", "aliases": [], "description": "The first Claude available without a sales conversation, and the model that introduced most developers to Anthropic. It used the Human/Assistant text format rather than structured messages — a shape that has not existed in the API since. Retired July 2025.", "released_on": "2023-07-11", "deprecated_on": "2025-01-21", "shutdown_on": "2025-07-21", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-opus-4-8", "recommended": true, "note": "This is an API migration, not a model swap: prompts move from a single Human/Assistant string to a messages array with a separate system field.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-2.1", "name": "Claude 2.1", "aliases": [], "description": "The model that took Claude to a 200k context window in late 2023, when the rest of the field was at 8k to 32k. Long context was Anthropic's clearest differentiator for most of a year. Retired July 2025.", "released_on": "2023-11-21", "deprecated_on": "2025-01-21", "shutdown_on": "2025-07-21", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-opus-4-8", "recommended": true, "note": "Claude 2.1 predates the Messages API. Migrating means moving off the legacy completions endpoint and its Human/Assistant prompt format entirely.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-3-5-haiku-20241022", "name": "Claude Haiku 3.5", "aliases": [ "claude-3-5-haiku-latest" ], "description": "The fast, cheap Claude that handled classification and routing work for most of 2025, and the last Haiku before the 4.5 generation. Retired February 2026.", "released_on": "2024-10-22", "deprecated_on": "2025-12-19", "shutdown_on": "2026-02-19", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Haiku 4.5 is substantially more capable at a similar price point, and supports extended thinking, which Haiku 3.5 did not.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-3-5-sonnet-20240620", "name": "Claude Sonnet 3.5", "aliases": [], "description": "The original Sonnet 3.5, from June 2024, and the model that introduced Artifacts. Anthropic later shipped a second model under the same 3.5 name, which is why anyone pinned to a date needs to check which one they meant. Retired October 2025.", "released_on": "2024-06-20", "deprecated_on": "2025-08-13", "shutdown_on": "2025-10-28", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-sonnet-4-6", "recommended": true, "note": "Anthropic points both Sonnet 3.5 snapshots at the same successor, so the June/October distinction stops mattering after the migration.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-3-5-sonnet-20241022", "name": "Claude Sonnet 3.5 v2", "aliases": [ "claude-3-5-sonnet-latest" ], "description": "The October 2024 Sonnet — confusingly the second model to carry the 3.5 name, which is why the dated id matters here more than anywhere else in the Claude lineup. It introduced computer use and was, for months, the model developers preferred over anything else on the market. Retired October 2025.", "released_on": "2024-10-22", "deprecated_on": "2025-08-13", "shutdown_on": "2025-10-28", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-sonnet-4-6", "recommended": true, "note": "The computer-use tool this model introduced has changed shape since; re-read the tool docs rather than porting the 2024 schema.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-3-7-sonnet-20250219", "name": "Claude Sonnet 3.7", "aliases": [ "claude-3-7-sonnet-latest" ], "description": "The first Claude with extended thinking, and the release that made a visible reasoning budget a mainstream API feature. It was retired in February 2026, almost exactly a year after launch.", "released_on": "2025-02-19", "deprecated_on": "2025-10-28", "shutdown_on": "2026-02-19", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-sonnet-4-6", "recommended": true, "note": "Sonnet 4.6 uses adaptive thinking rather than an explicit budget, so the thinking.budget_tokens field Sonnet 3.7 introduced no longer applies.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-3-haiku-20240307", "name": "Claude Haiku 3", "aliases": [], "description": "The cheapest Claude ever shipped, and the reason a lot of high-volume pipelines chose Anthropic at all. It outlived Haiku 3.5 by two months — Anthropic retired the newer model first, which is rarer than it sounds and caught out anyone who had migrated forward.", "released_on": "2024-03-07", "deprecated_on": "2026-02-19", "shutdown_on": "2026-04-20", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Haiku 4.5 costs more per token than Haiku 3 did. For pipelines chosen on price alone, re-run the unit economics before assuming this is a drop-in.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-3-opus-20240229", "name": "Claude Opus 3", "aliases": [], "description": "The first Claude to beat GPT-4 on published benchmarks, and the model that put Anthropic in serious contention. It had an unusually long deprecation window — six months — and its retirement prompted Anthropic's public commitments on preserving model weights after shutdown.", "released_on": "2024-02-29", "deprecated_on": "2025-06-30", "shutdown_on": "2026-01-05", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-opus-4-8", "recommended": true, "note": "Opus 4.8 is cheaper per token than Opus 3 was, has a 1M-token context window against 200k, and rejects temperature, top_p and top_k.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-3-sonnet-20240229", "name": "Claude Sonnet 3", "aliases": [], "description": "The middle model of the original Claude 3 family, positioned between Haiku and Opus on the price ladder that every provider has copied since. Retired July 2025.", "released_on": "2024-02-29", "deprecated_on": "2025-01-21", "shutdown_on": "2025-07-21", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-sonnet-4-6", "recommended": true, "note": "Two full generations separate these models. Prompts written to compensate for Sonnet 3's weaknesses usually hurt rather than help on 4.6.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-fable-5", "name": "Claude Fable 5", "aliases": [], "description": "Anthropic's most capable widely released model, built for agents that run for hours rather than turns. Anthropic publishes a \"not sooner than\" date for every active model, which is unusual candour: you know the floor on a model's life the day it ships.", "released_on": "2026-06-09", "earliest_shutdown_on": "2027-06-09", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "name": "Claude Haiku 4.5", "aliases": [ "claude-haiku-4-5" ], "description": "The fastest Claude with near-frontier quality, and the named replacement for every retired Haiku and Instant model going back to 2024. Active, with an earliest retirement date of October 2026.", "released_on": "2025-10-01", "earliest_shutdown_on": "2026-10-15", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-instant-1.0", "name": "Claude Instant 1.0", "aliases": [], "description": "The first of Anthropic's cheap, fast models — the ancestor of everything now called Haiku. Instant traded depth for latency and price, which is what made Claude viable for classification and moderation work in 2023.", "deprecated_on": "2024-09-04", "shutdown_on": "2024-11-06", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Two generations and a whole API separate these. Claude 1 predates the Messages API, so a migration means moving off /v1/complete and its \"\\n\\nHuman:\"/\"\\n\\nAssistant:\" prompt format entirely.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-instant-1.1", "name": "Claude Instant 1.1", "aliases": [], "description": "The middle Instant release, and the one that widened the context window enough to make document summarisation practical at Instant pricing. Retired alongside the whole Claude 1 generation in November 2024.", "deprecated_on": "2024-09-04", "shutdown_on": "2024-11-06", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Two generations and a whole API separate these. Claude 1 predates the Messages API, so a migration means moving off /v1/complete and its \"\\n\\nHuman:\"/\"\\n\\nAssistant:\" prompt format entirely.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-instant-1.2", "name": "Claude Instant 1.2", "aliases": [], "description": "The last of the Instant line — Anthropic's cheap, fast tier before Haiku existed. It was retired in November 2024 with two months notice, and the Haiku models inherited its role.", "released_on": "2023-08-09", "deprecated_on": "2024-09-04", "shutdown_on": "2024-11-06", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-haiku-4-5-20251001", "recommended": true, "note": "Haiku 4.5 is the modern equivalent of this tier. Like Claude 2, Instant predates the Messages API, so the request format changes too.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-mythos-5", "name": "Claude Mythos 5", "aliases": [], "description": "The same underlying model as Claude Fable 5 with the cybersecurity and biology safeguards lifted, offered only to approved Project Glasswing customers. It shares Fable 5's specs and pricing and its own 1M context window, but it is the one current Claude id you cannot reach by signing up - access runs through an Anthropic, AWS or Google Cloud account team.", "released_on": "2026-06-09", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/models/overview", "title": "Anthropic models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-mythos-preview", "name": "Claude Mythos Preview", "aliases": [], "description": "The first Mythos-class model, released in April 2026 to a closed group of cyber defenders and critical infrastructure providers under Project Glasswing. It is the only deprecated Claude model with no retirement date on the deprecations page - the notice lives in a callout on the models overview instead, because an invitation-only model has no public user base to give sixty days' notice to.", "status": "deprecated", "replacements": [ { "provider": "anthropic", "model": "claude-mythos-5", "recommended": true, "note": "Same access model: invitation-only through Project Glasswing, no self-serve sign-up. Mythos 5 is priced at less than half of Mythos Preview.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-08-03" }, { "url": "https://platform.claude.com/docs/en/about-claude/models/overview", "title": "Anthropic models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "anthropic", "model": "claude-opus-4-1-20250805", "name": "Claude Opus 4.1", "aliases": [ "claude-opus-4-1" ], "description": "The August 2025 Opus, and the last Anthropic model priced at $15/$75 per million tokens — three times what Opus costs today. It is the only deprecated Anthropic model still answering requests, and it has days left.", "released_on": "2025-08-05", "deprecated_on": "2026-06-05", "shutdown_on": "2026-08-05", "status": "deprecated", "replacements": [ { "provider": "anthropic", "model": "claude-opus-4-8", "recommended": true, "note": "Opus 4.7 and later reject temperature, top_p and top_k with a 400 error when set to a non-default value. Remove them from the request rather than tuning them down. Opus 4.8 also defaults effort to high on every surface, so set it explicitly if you were relying on a cheaper default.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "anthropic", "model": "claude-opus-4-20250514", "name": "Claude Opus 4", "aliases": [ "claude-opus-4-0" ], "description": "The Opus that launched the Claude 4 generation and introduced extended thinking with tool use. It was retired in June 2026 on two months notice, alongside Sonnet 4 — the first time Anthropic retired an entire generation in one announcement.", "released_on": "2025-05-14", "deprecated_on": "2026-04-14", "shutdown_on": "2026-06-15", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-opus-4-8", "recommended": true, "note": "Opus 4.7 and later reject temperature, top_p and top_k with a 400 error when set to a non-default value. Remove them from the request rather than tuning them down. Extended thinking has also been replaced by adaptive thinking: thinking.type: enabled is no longer accepted on Opus 4.7 and later.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-opus-4-5-20251101", "name": "Claude Opus 4.5", "aliases": [ "claude-opus-4-5" ], "description": "The November 2025 Opus, and the oldest Opus still active. Its earliest retirement date of November 2026 is the nearest of any active Anthropic model — worth noting if it is still the default in your config.", "released_on": "2025-11-01", "earliest_shutdown_on": "2026-11-24", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-opus-4-6", "name": "Claude Opus 4.6", "aliases": [], "description": "The first Claude to use the dateless id format, where claude-opus-4-6 is itself a pinned snapshot rather than a pointer to a dated one. It is active, with an earliest retirement of February 2027, and still accepts the sampling parameters that later Opus releases reject.", "earliest_shutdown_on": "2027-02-05", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-opus-4-7", "name": "Claude Opus 4.7", "aliases": [], "description": "The release where Anthropic deprecated the sampling parameters themselves: temperature, top_p and top_k return a 400 error on this model and everything after it. That is a deprecation you hit at runtime rather than on a schedule, and it breaks more code than most model retirements do.", "earliest_shutdown_on": "2027-04-16", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-opus-4-8", "name": "Claude Opus 4.8", "aliases": [], "description": "The last Opus of the 4.x line, and the successor Anthropic names for Opus 4.1, Opus 4, Opus 3 and the Claude 2 family — four generations of migration paths converge here. Still active, with an earliest retirement date of May 2027.", "earliest_shutdown_on": "2027-05-28", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-opus-5", "name": "Claude Opus 5", "aliases": [], "description": "The recommended default for complex agentic coding and enterprise work, and the most common migration target on the Anthropic side of this catalog. Its earliest possible retirement is July 2027 — a longer guaranteed floor than any other model here.", "earliest_shutdown_on": "2027-07-24", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-sonnet-4-20250514", "name": "Claude Sonnet 4", "aliases": [ "claude-sonnet-4-0" ], "description": "The workhorse of the Claude 4 generation, and for a year the default in Claude Code. Retired June 15, 2026 in the same announcement as Opus 4.", "released_on": "2025-05-14", "deprecated_on": "2026-04-14", "shutdown_on": "2026-06-15", "status": "retired", "replacements": [ { "provider": "anthropic", "model": "claude-sonnet-4-6", "recommended": true, "note": "Sonnet 4.6 keeps the same pricing and raises the context window to 1M tokens. Its adaptive thinking replaces the explicit thinking budget Sonnet 4 used, so any code setting thinking.budget_tokens needs review.", "external": false } ], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "anthropic", "model": "claude-sonnet-4-5-20250929", "name": "Claude Sonnet 4.5", "aliases": [ "claude-sonnet-4-5" ], "description": "The September 2025 Sonnet, still active but with the second-nearest earliest retirement date of any live Anthropic model. It is the last Sonnet to use explicit extended thinking rather than the adaptive thinking introduced in 4.6.", "released_on": "2025-09-29", "earliest_shutdown_on": "2026-09-29", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-sonnet-4-6", "name": "Claude Sonnet 4.6", "aliases": [], "description": "The Sonnet that replaced both Sonnet 4 and Sonnet 3.7 when those were retired, which makes it the busiest migration target in the Sonnet line. Active, with an earliest retirement of February 2027.", "earliest_shutdown_on": "2027-02-17", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "anthropic", "model": "claude-sonnet-5", "name": "Claude Sonnet 5", "aliases": [], "description": "The balanced tier of the Claude 5 generation, and the one carrying introductory pricing through August 2026. Not deprecated, with an earliest retirement date of June 2027.", "earliest_shutdown_on": "2027-06-30", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.claude.com/docs/en/about-claude/model-deprecations", "title": "Anthropic model deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "cohere", "model": "c4ai-aya-expanse-32b", "name": "Aya Expanse 32B", "aliases": [], "description": "The 32B multilingual research model covering 23 languages, and the survivor of the Aya Expanse pair - the 8B version was retired in April 2026 with no notice period at all. Nothing in Cohere's docs says this one is next, but it is the last Expanse standing.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "c4ai-aya-expanse-8b", "name": "Aya Expanse 8B", "aliases": [], "description": "The small half of Cohere Labs' 23-language Aya Expanse release, and a model a lot of people picked precisely because 8B was the size they could self-host. It was retired on 4 April 2026, the same day it was announced. The 32B version survives; this one does not.", "deprecated_on": "2026-04-04", "shutdown_on": "2026-04-04", "status": "retired", "replacements": [ { "provider": "cohere", "model": "command-a-03-2025", "recommended": true, "note": "Cohere names command-a-03-2025 and command-a-reasoning-08-2025, both far larger than 8B. If size was the point, command-r7b-12-2024 is the nearest supported model.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "cohere", "model": "c4ai-aya-vision-32b", "name": "Aya Vision 32B", "aliases": [], "description": "Cohere Labs' multimodal research model, 32B parameters and a 16k window, capable of answering questions about images across 23 languages. Its 8B sibling is already shut down, which makes this the only Aya model that still takes an image at all.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "c4ai-aya-vision-8b", "name": "Aya Vision 8B", "aliases": [], "description": "The compact multilingual vision model from Cohere Labs, answering questions about images in 23 languages at a size you could actually deploy. Retired 4 April 2026 with no notice, and the replacements Cohere names for it are text models - the only vision-capable Aya left is the 32B.", "deprecated_on": "2026-04-04", "shutdown_on": "2026-04-04", "status": "retired", "replacements": [ { "provider": "cohere", "model": "c4ai-aya-vision-32b", "recommended": false, "note": "Cohere's notice points at command-a-03-2025 and command-a-reasoning-08-2025, neither of which takes images. If you need vision, the 32B Aya or command-a-vision-07-2025 are the real options.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "cohere", "model": "command", "name": "Command", "aliases": [], "description": "The original Cohere instruction model, an undated floating id from the generate-endpoint era that predates the whole Command R line. It was deprecated on 15 September 2025 in the same sweep that removed /v1/generate, /v1/classify and fine-tuning for it - the announcement was less a model retirement than the end of Cohere's first API generation.", "deprecated_on": "2025-09-15", "status": "deprecated", "replacements": [ { "provider": "cohere", "model": "command-a-03-2025", "recommended": true, "note": "If you were calling this through /v1/generate rather than /v1/chat, the endpoint is gone too. This is an API migration, not a model swap.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "cohere", "model": "command-a-03-2025", "name": "Command A", "aliases": [], "description": "The 111B model that reset Cohere's enterprise line in March 2025, built to run on two GPUs and be deployed privately, which is the thing Cohere's customers actually buy. It is the named replacement for almost every retired Command model, so it sits at the end of most migration chains on this provider.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "command-a-plus-05-2026", "name": "Command A Plus", "aliases": [], "description": "Cohere's current flagship, a mixture-of-experts model that folds vision, reasoning and translation into one id after two years of shipping those as three separate Command A variants. The 128k window is smaller than the 256k Command A it sits above, which is the trade for the extra capability.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "command-a-reasoning-08-2025", "name": "Command A Reasoning", "aliases": [], "description": "Cohere's first model with an extended thinking mode, released August 2025 with the same 256k window as Command A. It is one of the two successors Cohere names for the retired Aya Expanse and Aya Vision 8B models, which is a large jump in size for anyone taking that advice literally.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "command-a-translate-08-2025", "name": "Command A Translate", "aliases": [], "description": "A translation-specialised Command A covering 23 languages, and the only model in Cohere's range with an 8k context window - a deliberate cap, since translation requests are short and the throughput matters more than the window.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "command-a-vision-07-2025", "name": "Command A Vision", "aliases": [], "description": "The image-capable Command A, aimed squarely at enterprise document work: charts, scanned forms, OCR-style question answering. Cohere shipped it as a separate id rather than adding vision to the base model, a split that Command A Plus has since closed.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "command-light", "name": "Command Light", "aliases": [], "description": "The cheap, small sibling of the original Command, and for a long time the budget option for classification and short generation on Cohere. Deprecated 15 September 2025 with no shutdown date. Cohere named no small replacement for it in that notice - command-r7b-12-2024 is the closest thing in the current range.", "deprecated_on": "2025-09-15", "status": "deprecated", "replacements": [ { "provider": "cohere", "model": "command-r-08-2024", "recommended": true, "note": "Cohere's named alternatives are all considerably larger than command-light was. If cost per token was why you were on this id, command-r7b-12-2024 is the closer match even though the notice does not mention it.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "cohere", "model": "command-r-03-2024", "name": "Command R (03-2024)", "aliases": [ "command-r" ], "description": "The model that put Cohere on the map for retrieval-augmented generation, with citations and connectors built into the chat API rather than bolted on. Cohere deprecated it on 15 September 2025 and has not published a shutdown date. Under Cohere's own policy the date is assigned at deprecation time; here it was not, so the id is closed to new customers and still answering for existing ones.", "deprecated_on": "2025-09-15", "status": "deprecated", "replacements": [ { "provider": "cohere", "model": "command-r-08-2024", "recommended": true, "note": "Cohere also names command-r-plus-08-2024 and command-a-03-2025. Whichever you pick, the connectors parameter went away in the same announcement, so a RAG integration built on managed connectors needs rewriting rather than repointing.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "cohere", "model": "command-r-08-2024", "name": "Command R (08-2024)", "aliases": [], "description": "The August 2024 refresh of Command R, and the model Cohere named as the replacement when it deprecated the original 03-2024 snapshot a year later. It also absorbed all of Cohere's fine-tuning: after the March 2025 change, every Command fine-tune is built on this base.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "command-r-plus-04-2024", "name": "Command R+ (04-2024)", "aliases": [ "command-r-plus" ], "description": "Cohere's April 2024 flagship, the first open-weights model to seriously compete on multi-step tool use, and the one that made the undated command-r-plus alias worth pinning. Deprecated 15 September 2025 with no shutdown date published, which means existing users keep access indefinitely and new ones get none.", "deprecated_on": "2025-09-15", "status": "deprecated", "replacements": [ { "provider": "cohere", "model": "command-r-plus-08-2024", "recommended": true, "note": "The nearest snapshot, but Cohere lists command-a-03-2025 as the strongest-performing option across domains and that is where new work should go.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "cohere", "model": "command-r-plus-08-2024", "name": "Command R+ (08-2024)", "aliases": [], "description": "The larger half of the August 2024 Command R refresh, tuned for multi-step RAG and tool use. It is deprecated-adjacent rather than deprecated: Cohere names it as a migration target for the retired 04-2024 snapshot while Command A has otherwise taken over the top of the range.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "command-r7b-12-2024", "name": "Command R7B", "aliases": [], "description": "A 7B model with a 128k window, the smallest thing in Cohere's Command line and the one aimed at running on a single commodity GPU. Cohere names it first among the chat replacements for the retired Aya 8B models, which makes it the like-for-like size match in that migration.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "embed-english-light-v2.0", "name": "Embed English Light v2", "aliases": [], "description": "The 1024-dimension light English embedder, chosen by people who cared about index size more than recall. Retired on 4 April 2026 with no notice period. Note that its direct successor embed-english-light-v3.0 is still active, so this is one of the few v2.0 retirements with a genuine like-for-like target.", "deprecated_on": "2026-04-04", "shutdown_on": "2026-04-04", "status": "retired", "replacements": [ { "provider": "cohere", "model": "embed-english-light-v3.0", "recommended": false, "note": "Cohere's notice names embed-english-v3.0 and embed-v4.0, neither of which is a light model. The light v3.0 id is the closer match on vector size and cost, but it is our recommendation rather than Cohere's.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "cohere", "model": "embed-english-light-v3.0", "name": "Embed English Light v3", "aliases": [], "description": "The small, fast English embedder - fewer dimensions, lower latency, cheaper at volume. Cohere kept the light variants alive through the v2.0 retirement even though embed-v4.0 has no light counterpart, which makes this the only route to a compact Cohere embedding today.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "embed-english-v2.0", "name": "Embed English v2", "aliases": [], "description": "Cohere's 4096-dimension English embedder from 2022, and one of the first embedding APIs anybody built a vector database on. It was retired effective 4 April 2026 - the same date as the announcement, so there was no notice period at all, which for an embedding model means re-indexing your entire corpus with no warning.", "deprecated_on": "2026-04-04", "shutdown_on": "2026-04-04", "status": "retired", "replacements": [ { "provider": "cohere", "model": "embed-v4.0", "recommended": true, "note": "Cohere also names embed-english-v3.0. Either way the vectors are not comparable with v2.0's, so this is a full re-embed of every document you have stored, not a config change.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "cohere", "model": "embed-english-v3.0", "name": "Embed English v3", "aliases": [], "description": "The English-only v3 embedder, and the model Cohere named when it retired embed-english-v2.0 in April 2026. It is still supported but no longer the default recommendation: new work is pointed at embed-v4.0, which handles images and much longer inputs.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "embed-multilingual-light-v3.0", "name": "Embed Multilingual Light v3", "aliases": [], "description": "The compact multilingual embedder, 512 tokens in and a smaller vector out. Like the English light model it survived the v2.0 cull, and it has no successor in the v4 generation - if you need small vectors across languages, this id is the whole option set.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "embed-multilingual-v2.0", "name": "Embed Multilingual v2", "aliases": [], "description": "The model that made cross-language retrieval practical for most teams - one vector space across 100-plus languages, years before that was table stakes. It was switched off on 4 April 2026 on the day it was announced, ending Cohere's entire v2.0 embedding generation in a single notice.", "deprecated_on": "2026-04-04", "shutdown_on": "2026-04-04", "status": "retired", "replacements": [ { "provider": "cohere", "model": "embed-v4.0", "recommended": true, "note": "embed-multilingual-v3.0 is the other option Cohere names. v4.0 takes 128k of input against v2.0's 512 tokens, so chunking strategy is worth revisiting at the same time as the re-index.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "cohere", "model": "embed-multilingual-v3.0", "name": "Embed Multilingual v3", "aliases": [], "description": "The 100-plus language embedder that replaced embed-multilingual-v2.0 when that model was retired in April 2026. Cross-lingual retrieval was Cohere's strongest differentiator for years, and this is the id most of that work was built on.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "embed-v4.0", "name": "Embed v4", "aliases": [], "description": "Cohere's current embedding model, and the first that takes text, images and mixed documents in one call with a configurable output dimension. The 128k input window is unusual for an embedder - the v3.0 family it replaces caps at 512 tokens, a 250x difference that changes how you chunk.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "rerank-english-v2.0", "name": "Rerank English v2", "aliases": [], "description": "The English reranker that made two-stage retrieval a standard pattern: cheap vector search first, then this model to reorder the top fifty. It is the one Cohere deprecation with a textbook runway - announced 2 December 2024, shut down 30 April 2025, five months apart.", "deprecated_on": "2024-12-02", "shutdown_on": "2025-04-30", "status": "retired", "replacements": [ { "provider": "cohere", "model": "rerank-v3.5", "recommended": true, "note": "v3.5 is multilingual, so the English-only and multilingual v2.0 ids collapse into one. Fine-tunes built on the v2.0 bases were explicitly not affected by this deprecation.", "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "cohere", "model": "rerank-english-v3.0", "name": "Rerank English v3", "aliases": [], "description": "The English reranker from the v3 generation, handling documents and JSON records alike with a 4k window. It has outlived the v2.0 models it succeeded and the v3.5 model that superseded it is still active too, so Cohere is currently running four generations of reranker at once.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "rerank-multilingual-v2.0", "name": "Rerank Multilingual v2", "aliases": [], "description": "The cross-language half of the v2.0 rerank pair, retired alongside its English sibling on 30 April 2025. Its existence is the reason rerank-v3.5 was a simplification as much as an upgrade: from v3.5 onward you no longer choose a language scope when you pick a reranker.", "deprecated_on": "2024-12-02", "shutdown_on": "2025-04-30", "status": "retired", "replacements": [ { "provider": "cohere", "model": "rerank-v3.5", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.cohere.com/docs/deprecations", "title": "Cohere deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "cohere", "model": "rerank-multilingual-v3.0", "name": "Rerank Multilingual v3", "aliases": [], "description": "The non-English reranker of the v3 generation. Cohere stopped splitting rerank by language at v3.5, so this is the last id in the line where you had to pick English or multilingual up front rather than getting both from one model.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "rerank-v3.5", "name": "Rerank v3.5", "aliases": [], "description": "The December 2024 reranker that replaced the entire v2.0 family, and still the model Cohere's deprecation page names for anyone migrating off it. Superseded in practice by the v4 pair, but not itself deprecated - it remains the cheapest supported way to rerank.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "rerank-v4.0-fast", "name": "Rerank v4 Fast", "aliases": [], "description": "The latency-optimised half of the v4 rerank pair, same 32k window and a lower quality ceiling. Splitting rerank into fast and pro tiers is new in v4; every earlier generation shipped one model per language scope instead.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "rerank-v4.0-pro", "name": "Rerank v4 Pro", "aliases": [], "description": "The current high-quality reranker, multilingual with a 32k context - eight times what rerank-v3.5 accepts, which means whole documents rather than passages. Third generation of a product line where Cohere has consistently been the strongest option on the market.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "tiny-aya-earth", "name": "Tiny Aya Earth", "aliases": [], "description": "The Tiny Aya variant tuned for West Asian and African languages. Splitting a 3B model four ways by region is the opposite of the usual approach of one large multilingual model, and it is aimed at exactly the languages that a global model under-serves.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "tiny-aya-fire", "name": "Tiny Aya Fire", "aliases": [], "description": "The South Asian language specialist in the Tiny Aya set, 8k context and small enough to run on a laptop. Cohere Labs releases these as research artefacts, so they carry no enterprise support commitment - a distinction that matters given what happened to the Aya Expanse and Aya Vision 8B ids.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "tiny-aya-global", "name": "Tiny Aya Global", "aliases": [], "description": "A 3.35B research model from Cohere Labs covering 70 languages, and the general member of a four-model set where the other three each specialise by region. Aya has always been the research arm's line rather than the commercial one, which is also why 8B-era Aya ids were retired with no notice period.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "cohere", "model": "tiny-aya-water", "name": "Tiny Aya Water", "aliases": [], "description": "The European and Asia-Pacific member of the Tiny Aya family. Its language list overlaps heavily with the models Cohere already sells commercially, which makes it the one Tiny Aya where the research and product lines answer the same question.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.cohere.com/docs/models", "title": "Cohere models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "deepseek", "model": "deepseek-chat", "name": "DeepSeek Chat", "aliases": [], "description": "For two years this was the DeepSeek API: a single moving id that always pointed at whatever the current chat model happened to be. It carried V2 in May 2024, then V2.5, V3, V3.1 and V3.2, so code written in 2024 kept working through five model generations without an edit. V4 ended that. DeepSeek shipped explicit ids, gave the alias three months, and switched it off on 24 July 2026.", "deprecated_on": "2026-04-24", "shutdown_on": "2026-07-24", "status": "retired", "replacements": [ { "provider": "deepseek", "model": "deepseek-v4-flash", "recommended": true, "note": "Through the transition deepseek-chat routed to V4-Flash with thinking off. Swapping the id alone is not equivalent: on deepseek-v4-flash the `thinking.type` parameter defaults to `enabled`, so you get reasoning tokens, slower responses and a larger bill unless you send `\"thinking\": {\"type\": \"disabled\"}`.", "external": false } ], "sources": [ { "url": "https://api-docs.deepseek.com/news/news260424/", "title": "DeepSeek V4 Preview Release", "accessed": "2026-08-03" }, { "url": "https://api-docs.deepseek.com/updates/", "title": "DeepSeek API change log", "accessed": "2026-08-03" }, { "url": "https://api-docs.deepseek.com/api/create-chat-completion", "title": "DeepSeek Create Chat Completion", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "deepseek", "model": "deepseek-coder", "name": "DeepSeek Coder", "aliases": [], "description": "DeepSeek's code-specialised endpoint, last upgraded on its own in July 2024 with DeepSeek-Coder-V2-0724. On 5 September 2024 DeepSeek merged the coder and chat lines into V2.5 and kept this id alive purely so existing integrations would not break. It has named nothing of its own since. DeepSeek has never published a shutdown date for it, and it was not named in the July 2026 retirement notice that took down deepseek-chat and deepseek-reasoner.", "deprecated_on": "2024-09-05", "status": "deprecated", "replacements": [ { "provider": "deepseek", "model": "deepseek-v4-flash", "recommended": true, "note": "DeepSeek's instruction in 2024 was to call deepseek-chat instead, but that id was itself retired in July 2026, so the live end of the path is deepseek-v4-flash. Send `\"thinking\": {\"type\": \"disabled\"}` to keep the non-reasoning behaviour deepseek-coder had.", "external": false } ], "sources": [ { "url": "https://api-docs.deepseek.com/updates/", "title": "DeepSeek API change log", "accessed": "2026-08-03" }, { "url": "https://api-docs.deepseek.com/quick_start/pricing", "title": "DeepSeek models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "deepseek", "model": "deepseek-reasoner", "name": "DeepSeek Reasoner", "aliases": [], "description": "The id that served DeepSeek-R1 in January 2025, the open-weights reasoning model that made chain-of-thought cheap and briefly moved the US stock market. It was the first mainstream API to hand back the reasoning trace itself in `reasoning_content`. From V3.1 onwards it stopped being a separate model and became the thinking half of a hybrid one, which is why V4 replaced it with a request parameter rather than another id.", "released_on": "2025-01-20", "deprecated_on": "2026-04-24", "shutdown_on": "2026-07-24", "status": "retired", "replacements": [ { "provider": "deepseek", "model": "deepseek-v4-flash", "recommended": true, "note": "deepseek-reasoner routed to V4-Flash with thinking on, which is the default, so `model: \"deepseek-v4-flash\"` reproduces the old behaviour. Use `thinking.reasoning_effort` (`low`, `high`, `max`) to trade depth against cost, and deepseek-v4-pro for the hardest problems.", "external": false } ], "sources": [ { "url": "https://api-docs.deepseek.com/news/news250120/", "title": "DeepSeek-R1 Release", "accessed": "2026-08-03" }, { "url": "https://api-docs.deepseek.com/news/news260424/", "title": "DeepSeek V4 Preview Release", "accessed": "2026-08-03" }, { "url": "https://api-docs.deepseek.com/updates/", "title": "DeepSeek API change log", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "deepseek", "model": "deepseek-v4-flash", "name": "DeepSeek V4 Flash", "aliases": [], "description": "The smaller half of DeepSeek V4 — 284B parameters with 13B active — and the landing place for every integration that used to call deepseek-chat or deepseek-reasoner. Thinking is a request parameter here rather than a separate id, which is what made those two aliases redundant. A million tokens of context at fourteen cents per million input is the reason it is the default migration target on the DeepSeek side of this catalog.", "released_on": "2026-04-24", "status": "active", "replacements": [], "sources": [ { "url": "https://api-docs.deepseek.com/quick_start/pricing", "title": "DeepSeek models and pricing", "accessed": "2026-08-03" }, { "url": "https://api-docs.deepseek.com/news/news260424/", "title": "DeepSeek V4 Preview Release", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "deepseek", "model": "deepseek-v4-pro", "name": "DeepSeek V4 Pro", "aliases": [], "description": "DeepSeek's frontier model, 1.6T parameters with 49B active, released alongside V4 Flash in April 2026 and served from the same base URL through either the OpenAI or the Anthropic wire format. No legacy alias ever pointed at it: the two ids DeepSeek retired in July 2026 both routed to Flash, so reaching Pro has always meant naming it. Roughly three times the price of Flash for the same million-token context window.", "released_on": "2026-04-24", "status": "active", "replacements": [], "sources": [ { "url": "https://api-docs.deepseek.com/quick_start/pricing", "title": "DeepSeek models and pricing", "accessed": "2026-08-03" }, { "url": "https://api-docs.deepseek.com/news/news260424/", "title": "DeepSeek V4 Preview Release", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "google", "model": "gemini-1.0-flash", "name": "Gemini 1.0 Flash", "aliases": [], "description": "The low-latency tier of the first Gemini generation, and the start of the Flash naming that Google still uses. Like 1.0 Pro it left the API without a published retirement date, which is why this page carries a status but no shutdown day.", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-2.5-flash", "recommended": true, "note": "Flash has stayed the name for the fast tier across four generations, so this is the natural target — but check the current Flash id, since 2.5 Flash is itself scheduled for retirement.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" }, { "url": "https://github.com/techdevsynergy/llm-model-deprecation", "title": "llm-model-deprecation registry (third-party)", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-1.0-pro", "name": "Gemini 1.0 Pro", "aliases": [], "description": "The first Gemini available through an API, launched in December 2023 as Googles answer to GPT-4. It shipped before Google settled on the model naming and lifecycle conventions used since, and no shutdown date was ever published for it — the registry that tracks it records the model as superseded with no retirement date on file.", "released_on": "2023-12-13", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-2.5-pro", "recommended": true, "note": "Nothing about a 1.0-era integration survives unchanged: the SDK, the request shape and the safety-setting enums have all been replaced since.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" }, { "url": "https://github.com/techdevsynergy/llm-model-deprecation", "title": "llm-model-deprecation registry (third-party)", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-1.5-flash", "name": "Gemini 1.5 Flash", "aliases": [ "gemini-1.5-flash-002" ], "description": "The cheap, fast half of the 1.5 generation, and the model that pushed long-context pricing down across the whole market. As with 1.5 Pro, Google has removed it from its published deprecation tables, so the shutdown date here is sourced from the community llm-model-deprecation registry.", "released_on": "2024-05-14", "shutdown_on": "2026-06-17", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-2.5-flash", "recommended": true, "note": "2.5 Flash is the direct successor, but note it carries its own shutdown date of October 16, 2026 — migrating here buys about a year, not a decade.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" }, { "url": "https://github.com/techdevsynergy/llm-model-deprecation", "title": "llm-model-deprecation registry (third-party)", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-1.5-pro", "name": "Gemini 1.5 Pro", "aliases": [ "gemini-1.5-pro-002" ], "description": "The model that made a million-token context window a mainstream expectation, and for a year the reason teams chose Gemini at all. Google no longer lists it on its deprecations page — rows are removed once a model is long gone — so the shutdown date here comes from the community llm-model-deprecation registry rather than a Google page, and should be treated as less firm than the rest of this site.", "released_on": "2024-02-15", "shutdown_on": "2026-06-17", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-2.5-pro", "recommended": true, "note": "The 2.5 generation thinks before answering by default, which changes both latency and token cost. Budget with thinkingConfig rather than assuming 1.5-era numbers carry over.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" }, { "url": "https://github.com/techdevsynergy/llm-model-deprecation", "title": "llm-model-deprecation registry (third-party)", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.0-flash", "name": "Gemini 2.0 Flash", "aliases": [ "gemini-2.0-flash-001" ], "description": "The model that made native tool use and a million-token context standard at Flash pricing, and the default Gemini for most of 2025. It was shut down on June 1, 2026 — the largest single removal of live traffic in Gemini's history, since so many integrations had never moved off it.", "released_on": "2025-02-05", "shutdown_on": "2026-06-01", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.6-flash", "recommended": true, "note": "Google skips the whole 2.5 and 3.x Flash lineage and names 3.6 Flash directly. Response formats are compatible; the thinking configuration is not, since 2.0 Flash had none.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.0-flash-lite", "name": "Gemini 2.0 Flash-Lite", "aliases": [ "gemini-2.0-flash-lite-001" ], "description": "The cheapest Gemini of the 2.0 generation, aimed at classification and extraction where per-call cost dominated. Shut down June 1, 2026 alongside gemini-2.0-flash.", "released_on": "2025-02-25", "shutdown_on": "2026-06-01", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.1-flash-lite", "recommended": true, "note": "The Lite tier persisted across generations, so this is a like-for-like swap — the closest thing to a drop-in in the Gemini catalog.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.0-flash-lite-preview", "name": "Gemini 2.0 Flash-Lite (preview)", "aliases": [ "gemini-2.0-flash-lite-preview-02-05" ], "description": "The preview of the 2.0 Lite tier, published on the same day as the 2.0 Flash launch. It outlived its own stable successor's preview cycle and was shut down in December 2025.", "released_on": "2025-02-05", "shutdown_on": "2025-12-09", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-2.5-flash-lite", "recommended": true, "note": "Google names 2.5 Flash-Lite here, which is itself scheduled for October 2026. gemini-3.1-flash-lite is the target that does not need migrating again.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.0-flash-live-001", "name": "Gemini 2.0 Flash Live", "aliases": [], "description": "The first Live API model, handling bidirectional audio and video streaming over a websocket rather than request-response. It shut down on 9 December 2025, six months before the rest of the 2.0 line, because the Live API moved fast enough that a 2.0-era session protocol was not worth maintaining.", "shutdown_on": "2025-12-09", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.1-flash-live-preview", "recommended": true, "note": "The successor is still a preview id. Migrating off a retired model onto a preview one is not ideal, but it is what Google names.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.5-flash", "name": "Gemini 2.5 Flash", "aliases": [], "description": "Probably the highest-volume Gemini model ever deployed: fast, cheap, and the first Flash with a configurable thinking budget. Active, but Google has published October 16, 2026 as the earliest date it could be retired.", "released_on": "2025-06-17", "earliest_shutdown_on": "2026-10-16", "status": "active", "replacements": [ { "provider": "google", "model": "gemini-3.6-flash", "recommended": true, "note": "Google skips 3.1 and 3.5 and names 3.6 Flash directly. The thinking-budget configuration introduced in 2.5 Flash has changed shape since — re-read the thinking docs rather than porting the config.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-2.5-flash-image", "name": "Gemini 2.5 Flash Image", "aliases": [], "description": "The original Nano Banana, and the model that made conversational image editing a mainstream feature in late 2025. Its shutdown is set two weeks earlier than the rest of the 2.5 line, on 2 October 2026 rather than 16 October, so an app doing both chat and image editing has two migration deadlines, not one.", "earliest_shutdown_on": "2026-10-02", "status": "active", "replacements": [ { "provider": "google", "model": "gemini-3.1-flash-image", "recommended": true, "note": "Google's deprecation table names gemini-3.1-flash-image-preview. That preview has since gone stable as gemini-3.1-flash-image, which is what the models page lists and what this page points at.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "google", "model": "gemini-2.5-flash-lite", "name": "Gemini 2.5 Flash-Lite", "aliases": [], "description": "The cheapest model of the 2.5 generation, built for classification and extraction at volume where the thinking budget could be set to zero. Active, with an earliest shutdown of October 16, 2026 alongside the rest of the 2.5 line.", "released_on": "2025-07-22", "earliest_shutdown_on": "2026-10-16", "status": "active", "replacements": [ { "provider": "google", "model": "gemini-3.1-flash-lite", "recommended": true, "note": "Note that Google points Flash-Lite at the 3.1 generation while pointing Flash at 3.6 — the Lite tiers are on their own release cadence.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-2.5-flash-lite-preview-09-2025", "name": "Gemini 2.5 Flash-Lite (preview 09-2025)", "aliases": [], "description": "A September 2025 preview of the Flash-Lite tier, notable mainly for its id format: it used a four-digit year where every other Gemini preview uses two. Shut down March 2026.", "released_on": "2025-09-25", "shutdown_on": "2026-03-31", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.1-flash-lite", "recommended": true, "note": "Google routes this straight to the 3.1 Lite tier rather than to the stable 2.5 Flash-Lite, which is itself scheduled for October 2026.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.5-flash-preview-05-20", "name": "Gemini 2.5 Flash (preview 05-20)", "aliases": [], "description": "The May 2025 preview of 2.5 Flash, and the first Gemini to expose a configurable thinking budget. Shut down November 2025, six months after release.", "released_on": "2025-05-20", "shutdown_on": "2025-11-18", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.6-flash", "recommended": true, "note": "The thinking-budget parameter this preview introduced has changed shape twice since; check the current thinking docs rather than porting config.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.5-flash-preview-09-25", "name": "Gemini 2.5 Flash (preview 09-25)", "aliases": [], "description": "The September 2025 preview snapshot of 2.5 Flash, released to test changes ahead of the stable model. Shut down February 2026.", "released_on": "2025-09-25", "shutdown_on": "2026-02-17", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.6-flash", "recommended": true, "note": "Google points this snapshot past the stable 2.5 Flash id straight to 3.6 Flash, so there is no reason to migrate to 2.5 Flash as an interim step.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.5-pro", "name": "Gemini 2.5 Pro", "aliases": [], "description": "The model that put Google back at the front of the field in mid-2025, with thinking enabled by default and a million-token context window. It is still active and widely deployed, but Google has published October 16, 2026 as its earliest shutdown — which for a model this widely used is the single most important date on this site.", "released_on": "2025-06-17", "earliest_shutdown_on": "2026-10-16", "status": "active", "replacements": [ { "provider": "google", "model": "gemini-3.1-pro-preview", "recommended": true, "note": "The successor still carries a -preview suffix, so a migration off 2.5 Pro currently means moving onto a preview id. Factor that into the timing.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-2.5-pro-preview-03-25", "name": "Gemini 2.5 Pro (preview 03-25)", "aliases": [], "description": "The first 2.5 Pro preview, from March 2025, and the release where Gemini started thinking before answering by default. It was widely benchmarked under this id, which is why it still appears in evaluation code. Shut down December 2025.", "released_on": "2025-03-25", "shutdown_on": "2025-12-02", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.1-pro-preview", "recommended": true, "note": "If you are reproducing a benchmark rather than shipping, note that no hosted model reproduces this snapshot's behaviour.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.5-pro-preview-05-06", "name": "Gemini 2.5 Pro (preview 05-06)", "aliases": [], "description": "The second 2.5 Pro preview, from May 2025. Google shipped three preview snapshots in as many months before the stable release, each a separate id, and retired all three together in December 2025.", "released_on": "2025-05-06", "shutdown_on": "2025-12-02", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.1-pro-preview", "recommended": true, "note": "Two generations separate these models. Treat this as a re-evaluation rather than a swap.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-2.5-pro-preview-06-05", "name": "Gemini 2.5 Pro (preview 06-05)", "aliases": [], "description": "The last of three 2.5 Pro previews, released twelve days before the stable model. All three were shut down on the same day in December 2025.", "released_on": "2025-06-05", "shutdown_on": "2025-12-02", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.1-pro-preview", "recommended": true, "note": "Google points every 2.5 Pro preview at 3.1 Pro rather than at stable 2.5 Pro, which is itself scheduled for retirement in October 2026.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-3-flash-preview", "name": "Gemini 3 Flash (preview)", "aliases": [], "description": "The first Flash model of the Gemini 3 generation. Google lists a recommended replacement for it but no shutdown date, which puts it in an unusual middle state: superseded but not yet scheduled.", "released_on": "2025-12-17", "status": "active", "replacements": [ { "provider": "google", "model": "gemini-3.6-flash", "recommended": true, "note": "Google names gemini-3.6-flash as the replacement even though no shutdown date has been set. Moving now avoids a rushed migration when one is.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-3-pro-image", "name": "Gemini 3 Pro Image", "aliases": [], "description": "Nano Banana Pro, the high-end image model of the Gemini 3 generation, released the same day as the 3.1 Flash Image. It is the only Pro-tier image model Google serves and it carries the Gemini 3 version number rather than 3.1, so the image line's numbering has already drifted from the text line's.", "released_on": "2026-05-28", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-08-03" }, { "url": "https://ai.google.dev/gemini-api/docs/models", "title": "Gemini API models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "google", "model": "gemini-3-pro-preview", "name": "Gemini 3 Pro (preview)", "aliases": [], "description": "The first Pro model of the Gemini 3 generation, and one of the shortest-lived entries in this catalog: released November 2025, shut down March 2026, less than four months later. Preview ids on the Gemini API carry real shutdown risk.", "released_on": "2025-11-18", "shutdown_on": "2026-03-09", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.1-pro-preview", "recommended": true, "note": "The successor is also a preview id. On the Gemini API, preview does not mean experimental so much as short-lived — plan to move again.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-3.1-flash-image", "name": "Gemini 3.1 Flash Image", "aliases": [], "description": "Nano Banana 2, the stable successor to the 2.5 Flash Image model, released May 2026. It reached a stable id faster than most Gemini image work - the previous generation spent months as a preview string that a lot of production code ended up pinned to.", "released_on": "2026-05-28", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-08-03" }, { "url": "https://ai.google.dev/gemini-api/docs/models", "title": "Gemini API models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "google", "model": "gemini-3.1-flash-lite", "name": "Gemini 3.1 Flash-Lite", "aliases": [], "description": "The Flash-Lite of the 3.1 generation, and the model Google names as the replacement for the 2.0 and 2.5 Lite tiers. It is not deprecated, but Google has already published May 2027 as the earliest date it could be retired.", "released_on": "2026-05-07", "earliest_shutdown_on": "2027-05-07", "status": "active", "replacements": [ { "provider": "google", "model": "gemini-3.5-flash-lite", "recommended": true, "note": "Named as the eventual successor. No action is needed yet — this model is current and fully supported.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-3.1-flash-lite-preview", "name": "Gemini 3.1 Flash-Lite (preview)", "aliases": [], "description": "The preview that became gemini-3.1-flash-lite. It ran for under three months before being shut down in favour of the stable id — the standard Gemini pattern, and the reason preview ids should never reach production config.", "released_on": "2026-03-03", "shutdown_on": "2026-05-25", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-3.1-flash-lite", "recommended": true, "note": "Dropping the \"-preview\" suffix is the entire migration. Behaviour is unchanged.", "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "google", "model": "gemini-3.1-flash-live-preview", "name": "Gemini 3.1 Flash Live Preview", "aliases": [], "description": "The current Live API model for realtime audio and video sessions, and the named successor to the retired 2.0 Live id. Google has never shipped a stable non-preview Live model: every generation of this line has carried a preview suffix, which means no committed lifecycle at all.", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/models", "title": "Gemini API models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "google", "model": "gemini-3.1-pro-preview", "name": "Gemini 3.1 Pro (preview)", "aliases": [], "description": "The current Pro-tier model, and the migration target Google names for gemini-2.5-pro and every retired 2.5 Pro preview. It has carried the \"-preview\" suffix since February 2026 without a stable id appearing — long enough that treating it as production is now the norm rather than the exception.", "released_on": "2026-02-19", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-3.5-flash", "name": "Gemini 3.5 Flash", "aliases": [], "description": "The Flash release between 3.1 and 3.6, still active with no shutdown date published. Gemini Flash models have historically run about sixteen months from release to shutdown, which is the useful planning number if you are pinning one.", "released_on": "2026-05-19", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-3.5-flash-lite", "name": "Gemini 3.5 Flash-Lite", "aliases": [], "description": "The cheapest tier of the current Gemini generation, and the named successor to gemini-3.1-flash-lite. No shutdown date published yet.", "released_on": "2026-07-21", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-3.6-flash", "name": "Gemini 3.6 Flash", "aliases": [], "description": "The current Flash model, and the successor Google names for the entire 2.x Flash line. Google publishes a shutdown date for almost every Gemini model as a matter of policy; this one has none yet, which is the clearest signal available that it is at the front of the queue rather than the back.", "released_on": "2026-07-21", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "google", "model": "gemini-embedding-001", "name": "Gemini Embedding 001", "aliases": [], "description": "The first Gemini-branded embedding model, and the holder of the longest runway in Google's entire deprecation table: released July 2025 with a shutdown no earlier than 14 May 2028, almost three years out. Google gives embedding models far more notice than chat models precisely because re-indexing is expensive.", "released_on": "2025-07-14", "earliest_shutdown_on": "2028-05-14", "status": "active", "replacements": [ { "provider": "google", "model": "gemini-embedding-2", "recommended": true, "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "google", "model": "gemini-embedding-2", "name": "Gemini Embedding 2", "aliases": [], "description": "The current Gemini embedding model, and the named successor to both gemini-embedding-001 and the older text-embedding-004. It is the one id every retired Google embedding chain terminates on, which makes it the safest place to point a new retrieval stack.", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/models", "title": "Gemini API models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "google", "model": "imagen-3.0-generate-002", "name": "Imagen 3", "aliases": [], "description": "Google's third-generation image model on the Gemini API, and the last Imagen to serve before the 4.0 line took over. It shut down on 10 November 2025. Imagen sits on a separate version track from Gemini itself, so nothing about a Gemini model's lifecycle tells you anything about the image models.", "shutdown_on": "2025-11-10", "status": "retired", "replacements": [ { "provider": "google", "model": "imagen-4.0-generate-001", "recommended": true, "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "google", "model": "imagen-4.0-generate-001", "name": "Imagen 4", "aliases": [], "description": "The current Imagen generation and the named replacement for Imagen 3. Google now runs two parallel image lines - Imagen for dedicated text-to-image, and the Gemini Flash Image models for conversational editing - and only the second gets the Nano Banana marketing.", "status": "active", "replacements": [], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "google", "model": "text-embedding-004", "name": "Text Embedding 004", "aliases": [], "description": "Google's pre-Gemini-branded embedding model from April 2024, and the id most early Vertex and AI Studio RAG stacks were built on. It shut down on 14 January 2026 after a twenty-one month run. Embedding retirements are the expensive kind: the vectors it produced are not comparable with any successor's, so migrating means re-embedding the whole corpus.", "released_on": "2024-04-09", "shutdown_on": "2026-01-14", "status": "retired", "replacements": [ { "provider": "google", "model": "gemini-embedding-2", "recommended": true, "external": false } ], "sources": [ { "url": "https://ai.google.dev/gemini-api/docs/deprecations", "title": "Gemini API deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "minimax", "model": "image-01", "name": "Image 01", "aliases": [], "description": "MiniMax's only image generation model, released February 2025 and never versioned since. Every other MiniMax product line has shipped three or four generations in that time, which makes this the one id where a replacement landing without warning would be a genuine surprise.", "released_on": "2025-02-15", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-H3", "name": "MiniMax H3", "aliases": [], "description": "The newest Hailuo video model, released July 2026 as a general-purpose multimodal video generator. Video is the product MiniMax is best known for outside developer circles, and the H-series numbering runs independently of the M-series language models.", "released_on": "2026-07-31", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-M2", "name": "MiniMax M2", "aliases": [], "description": "The October 2025 model that MiniMax released open-weights under an efficient model for the agentic era framing, and the one that got the company taken seriously outside China. It is the oldest language model still on the API, nine months and four generations old, with no retirement date published.", "released_on": "2025-10-27", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-M2.1", "name": "MiniMax M2.1", "aliases": [], "description": "A December 2025 release aimed at programming and large-scale code refactoring, and the first MiniMax model to ship with a highspeed twin. Eight months and three generations later it is still in the model list at its original context length.", "released_on": "2025-12-22", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-M2.1-highspeed", "name": "MiniMax M2.1 Highspeed", "aliases": [], "description": "The original highspeed variant, which established a naming convention MiniMax has applied to every model since. It is the oldest id in the pattern and, on a platform with no deprecation policy, the one with the least idea of how long it has left.", "released_on": "2025-12-22", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-M2.5", "name": "MiniMax M2.5", "aliases": [], "description": "The February 2026 model that MiniMax claimed state-of-the-art results with on programming, tool calling and search. It reads as the legacy option today - two generations back and still listed at full price - but MiniMax has published no deprecation notice for it, so it stays active in this catalogue.", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-M2.5-highspeed", "name": "MiniMax M2.5 Highspeed", "aliases": [], "description": "The fast-serving M2.5. Its continued presence in the model list is the clearest illustration of MiniMax's approach to lifecycle: ids accumulate and nothing is ever announced as going away, which is comfortable until the day something quietly stops answering.", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-M2.7", "name": "MiniMax M2.7", "aliases": [], "description": "The March 2026 release that MiniMax launched under a recursive self-improvement banner, and still the model the platform recommends when M3's million-token window is more than you need. MiniMax has never deprecated a model, so M2.7 sits alongside four older siblings rather than replacing them.", "released_on": "2026-03-18", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-M2.7-highspeed", "name": "MiniMax M2.7 Highspeed", "aliases": [], "description": "The throughput-optimised M2.7, sold as a distinct id rather than as a routing hint. MiniMax has shipped a highspeed twin for every model since M2.1, which doubles the size of the catalogue and means a latency decision is baked into the model string your code sends.", "released_on": "2026-03-18", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "MiniMax-M3", "name": "MiniMax M3", "aliases": [], "description": "MiniMax's current flagship, released June 2026 with a million-token window - five times any M2-series model - and built for agentic reasoning, tool use and multimodal chat input. It is the first MiniMax id to break the 204,800-token ceiling every model since M2 had shared.", "released_on": "2026-06-01", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "music-3.0", "name": "Music 3.0", "aliases": [], "description": "The July 2026 music generation model, fourth in a line that started as a beta in mid-2025 and has shipped roughly every three months since. Unlike the speech line, MiniMax lists only the current music model in its API overview, so the earlier ids are gone without ever having been formally retired.", "released_on": "2026-07-16", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" }, { "url": "https://platform.minimax.io/docs/api-reference/api-overview", "title": "MiniMax API models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "speech-02-hd", "name": "Speech 02 HD", "aliases": [], "description": "The April 2025 speech model, and the oldest audio id MiniMax still serves. Note the zero-padded version number: MiniMax switched from 02 to 2.5 without ever shipping an 03, so string-sorting the speech ids gives you the wrong order.", "released_on": "2025-04-02", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "speech-02-turbo", "name": "Speech 02 Turbo", "aliases": [], "description": "The low-latency member of the original Speech 02 pair from April 2025. Sixteen months on it is still in the API model list unchanged, which on this platform means nothing in particular - MiniMax lists models until it does not.", "released_on": "2025-04-02", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "speech-2.6-hd", "name": "Speech 2.6 HD", "aliases": [], "description": "The October 2025 high-fidelity speech model, superseded by 2.8 three months later but still listed and still priced. Speech models tend to outlive their replacements in practice because voice clones are tied to the model that produced them.", "released_on": "2025-10-29", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "speech-2.6-turbo", "name": "Speech 2.6 Turbo", "aliases": [], "description": "The fast variant of the 2.6 speech generation. Four MiniMax speech ids across three generations are currently active at once, which is a lot of surface area for a provider that has never published a deprecation.", "released_on": "2025-10-29", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "speech-2.8-hd", "name": "Speech 2.8 HD", "aliases": [], "description": "The high-fidelity half of MiniMax's January 2026 text-to-speech release, adding natural sound tags that let you mark laughter or hesitation inline in the text. HD trades latency for audio quality; the turbo variant makes the opposite trade.", "released_on": "2026-01-23", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "minimax", "model": "speech-2.8-turbo", "name": "Speech 2.8 Turbo", "aliases": [], "description": "The low-latency Speech 2.8, for voice agents where the pause before the model starts talking is the thing users notice. MiniMax has kept both an HD and a turbo id in every speech generation since 02, and has retired none of them.", "released_on": "2026-01-23", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.minimax.io/docs/release-notes/models", "title": "MiniMax model release notes", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "codestral-2405", "name": "Codestral 24.05", "aliases": [], "description": "The first Codestral, a 22B fill-in-the-middle model covering eighty programming languages, released under a licence that banned commercial use and started a long argument about what Mistral meant by open. It ran on the API for thirteen months.", "deprecated_on": "2024-12-02", "shutdown_on": "2025-06-16", "status": "retired", "replacements": [ { "provider": "mistral", "model": "codestral-25-08", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "codestral-25-08", "name": "Codestral 25.08", "aliases": [], "description": "The third Codestral to hold the name and the only one still answering. Mistral has kept a dedicated code-completion model through three generations while folding every other specialist line into Medium and Small, largely because fill-in-the-middle for IDE autocomplete is a different latency problem than chat. Both earlier Codestrals and Codestral Mamba point here.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "codestral-2501", "name": "Codestral 25.01", "aliases": [], "description": "The January 2025 Codestral, twice as fast as the 24.05 original and the version that got Codestral into Continue and Tabnine defaults. Note that Mistral's own table lists its replacement as \"Codestral\" without a version - the live id is codestral-25-08, and the undated codestral alias now points there.", "deprecated_on": "2025-11-06", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "codestral-25-08", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "devstral-2512", "name": "Devstral 2", "aliases": [], "description": "Mistral's agentic coding model, built for driving a codebase through a tool loop rather than completing a line in an editor. Devstral 2 lasted seven months. Mistral retired the whole Devstral line into Medium 3.5 in mid-2026, which is the clearest example of its strategy of collapsing specialist ids into one general model.", "deprecated_on": "2026-05-22", "shutdown_on": "2026-07-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "note": "Medium 3.5 is explicitly optimised for agentic and coding use, but it is a general model. Prompts tuned to Devstral's SWE-bench-style scaffolding are worth re-testing rather than lifting across unchanged.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "devstral-medium-2507", "name": "Devstral Medium 1.0", "aliases": [], "description": "The premier-tier half of the first Devstral generation, sold as the coding model you paid for when the open Devstral Small was not strong enough. It ran ten months before the February 2026 deprecation wave took it, and unlike Small it never had an open-weights release to fall back on.", "deprecated_on": "2026-02-27", "shutdown_on": "2026-05-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "devstral-small-2505", "name": "Devstral Small 1.0", "aliases": [], "description": "The first Devstral, released in May 2025 with All Hands AI as a model for OpenHands-style agents. It had the shortest deprecation runway of any Mistral model in the catalogue - announced 31 October 2025 and gone by 30 November, four weeks later - because 1.1 had already replaced it in practice.", "deprecated_on": "2025-10-31", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "devstral-small-2507", "name": "Devstral Small 1.1", "aliases": [], "description": "An Apache 2.0 coding model that punched well above its size on SWE-bench and got adopted fast by people running agents locally. The API id is retired, but because the weights were released openly this is one of the few entries in the catalogue where the model itself did not go anywhere - only Mistral's hosting of it did.", "deprecated_on": "2026-02-27", "shutdown_on": "2026-05-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "labs-devstral-small-2512", "name": "Devstral Small 2", "aliases": [], "description": "Shipped under the labs prefix, which on Mistral's API means exactly what it sounds like: no support commitment and no expectation of a long life. It got five weeks between deprecation and shutdown and never graduated to a stable id, unlike Leanstral, which made the same journey out of labs successfully.", "deprecated_on": "2026-02-27", "shutdown_on": "2026-03-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "labs-leanstral-2603", "name": "Leanstral 26.03", "aliases": [], "description": "The experimental id Leanstral shipped under before it was promoted. It was deprecated on 22 May 2026 and switched off on 30 June, which is the normal outcome for a labs id: the model survives under a new stable name and only the experimental address dies.", "deprecated_on": "2026-05-22", "shutdown_on": "2026-06-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "leanstral-1-5", "recommended": true, "note": "A rename rather than a swap, but the id changes shape completely - there is no date stamp on leanstral-1-5, so anything parsing the trailing YYMM out of the model string will break.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "labs-mistral-small-creative", "name": "Mistral Small Creative", "aliases": [], "description": "A writing-tuned Small that traded benchmark scores for prose that did not read like an assistant. It never left labs and lasted four months. Mistral pointed it at Ministral 3 8B rather than Mistral Small 4, which is a smaller model - a reminder that the named replacement is the provider's recommendation, not a claim of equivalence.", "deprecated_on": "2026-03-31", "shutdown_on": "2026-04-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "ministral-3-8b-25-12", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "leanstral-1-5", "name": "Leanstral 1.5", "aliases": [], "description": "Leanstral graduated out of Mistral's labs prefix with this release: the experimental id was labs-leanstral-2603 and it was retired five weeks after 1.5 landed. Losing the labs prefix is the signal that matters here, because ids under it carry no retirement guarantee at all and this one now sits in the supported Apache 2.0 range.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "magistral-medium-2506", "name": "Magistral Medium 1.0", "aliases": [], "description": "Mistral's first reasoning model, announced in June 2025 as Europe's answer to o1 and R1, with multilingual chains of thought as its distinguishing pitch. It lasted under five months on the API. Of everything in the Mistral table this is the shortest gap between a flagship launch and its deprecation notice.", "deprecated_on": "2025-10-31", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "magistral-medium-2507", "name": "Magistral Medium 1.1", "aliases": [], "description": "A July 2025 refresh that mostly cleaned up the reasoning traces and reduced the amount of language-switching mid-thought that plagued 1.0. It was deprecated three months later in the same October wave that took every other Magistral shipped that year, and had thirty days to live after the announcement.", "deprecated_on": "2025-10-31", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "magistral-medium-2509", "name": "Magistral Medium 1.2", "aliases": [], "description": "The last Magistral Medium, and the one that added vision to Mistral's reasoning line. It survived ten months, longer than either 1.0 or 1.1, and died with the rest of the family when Mistral folded explicit reasoning into Medium 3.5 as a mode rather than a model.", "deprecated_on": "2026-05-22", "shutdown_on": "2026-07-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "note": "Magistral exposed its chain of thought in a dedicated block. Medium 3.5 handles reasoning as a request-level setting, so the response shape is not the same and thinking-token accounting changes with it.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "magistral-small-2506", "name": "Magistral Small 1.0", "aliases": [], "description": "The first open reasoning model Mistral published, distilled from Magistral Medium and released under Apache 2.0 on the same June 2025 day. Its API life was under five months. If you are trying to reproduce a June 2025 evaluation run, the id no longer resolves and you will need the Hugging Face weights.", "deprecated_on": "2025-10-31", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "magistral-small-2507", "name": "Magistral Small 1.1", "aliases": [], "description": "The 24B open reasoning model from Mistral's July 2025 refresh, released alongside Magistral Medium 1.1 and deprecated on the same day three months later. It is one of four Magistrals that shared a 30 November 2025 shutdown, so the entire first year of Mistral's reasoning line went dark in one evening.", "deprecated_on": "2025-10-31", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "magistral-small-2509", "name": "Magistral Small 1.2", "aliases": [], "description": "The open-weights counterpart to Magistral Medium 1.2, Apache 2.0 and small enough to run on a single consumer GPU, which is why it showed up in a lot of local reasoning setups. Its API id retired on 31 July 2026 alongside the rest of the line, but the weights it shipped are still downloadable.", "deprecated_on": "2026-04-30", "shutdown_on": "2026-07-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "ministral-3-14b-25-12", "name": "Ministral 3 14B", "aliases": [], "description": "The largest of the three Ministral 3 edge models and the only one Mistral names as the successor to Pixtral 12B, which makes it the place multimodal small-model workloads ended up. Apache 2.0 like the rest of the Ministral 3 line, so unlike the 24.10 Ministrals it replaced there is nothing stopping you running it yourself when the API id eventually rotates.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "ministral-3-3b-25-12", "name": "Ministral 3 3B", "aliases": [], "description": "The smallest model Mistral serves over the API, and the successor to ministral-3b-2410. The 24.10 Ministrals were premier-licensed and lasted barely a year; the 25.12 rewrite is Apache 2.0 and shares its training run with the 8B and 14B, so the three now move as one family rather than as separately dated point releases.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "ministral-3-8b-25-12", "name": "Ministral 3 8B", "aliases": [], "description": "The middle Ministral 3, and the id that inherited the longest tail in Mistral's catalogue: open-mistral-7b, Mistral Nemo 12B, Ministral 8B and Mistral Small Creative all name it as their replacement. Anything you wrote against a seven-or-eight-billion-parameter Mistral endpoint between 2023 and 2026 eventually resolves here.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "ministral-3b-2410", "name": "Ministral 3B", "aliases": [], "description": "The smaller half of les Ministraux, and the only one of the pair whose weights Mistral never released in any form - it was API-only for its entire fourteen months. When it shut down at the end of 2025 the model became genuinely unavailable rather than merely unhosted.", "deprecated_on": "2025-12-02", "shutdown_on": "2025-12-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "ministral-3-3b-25-12", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "ministral-8b-2410", "name": "Ministral 8B", "aliases": [], "description": "Announced on Mistral 7B's first birthday as the edge-device successor to it, with an interleaved sliding-window attention scheme for cheap long context. Its research licence made it awkward to actually deploy on the edge, which the Apache-licensed Ministral 3 line later fixed.", "deprecated_on": "2025-12-02", "shutdown_on": "2025-12-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "ministral-3-8b-25-12", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-large-2402", "name": "Mistral Large 1.0", "aliases": [], "description": "The model Mistral launched with its Microsoft partnership in February 2024, available on Azure the same week it hit la Plateforme. It was the company's first closed-weights flagship and the moment the open-source-only framing stopped being accurate. Shut down 16 June 2025.", "deprecated_on": "2024-11-30", "shutdown_on": "2025-06-16", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-large-3-25-12", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-large-2407", "name": "Mistral Large 2.0", "aliases": [], "description": "The 123B model that made Mistral competitive with GPT-4 class systems in mid-2024, and the first Large released with downloadable weights under a research licence. It also holds the longest deprecation runway in the Mistral table: announced 30 November 2024, shut down 30 March 2025, four full months.", "deprecated_on": "2024-11-30", "shutdown_on": "2025-03-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-large-3-25-12", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-large-2411", "name": "Mistral Large 2.1", "aliases": [], "description": "For most of 2025 this was what mistral-large-latest resolved to, which means a lot of code that never named a version was running it. Mistral did not route it to Large 3 when it retired it - it pointed at Medium 3.5, which is the clearest statement that the Large tier stopped being the top of the range.", "deprecated_on": "2026-02-27", "shutdown_on": "2026-05-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "note": "Mistral names Medium 3.5, not Large 3, as the successor. If you want to stay in the Large tier for licensing reasons, mistral-large-3-25-12 exists and is Apache 2.0, but it is not the model Mistral points you at.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-large-3-25-12", "name": "Mistral Large 3", "aliases": [], "description": "Apache 2.0, which is the detail people miss: Mistral Large spent two generations as a closed premier model before Large 3 shipped open in December 2025. It is the named successor to both Mistral Large 1.0 and 2.0, and it is the only survivor of the Large line - Large 2.1 was pushed to Medium 3.5 instead, so the family's own name no longer marks the top of the range.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "mistral-medium-2312", "name": "Mistral Medium 1.0", "aliases": [], "description": "The oldest id in Mistral's table, from December 2023, and a model the company never published anything about - no weights, no paper, no architecture. It leaked to the internet as \"Miqu\" in early 2024 and was widely believed to be a Llama 2 70B fine-tune, which Mistral's CEO more or less confirmed.", "deprecated_on": "2024-11-30", "shutdown_on": "2025-06-16", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-medium-2505", "name": "Mistral Medium 3", "aliases": [], "description": "The model that reintroduced a middle tier to Mistral's API after two years of only Small and Large, and the first sign that Large was being quietly demoted. It shares its 31 August 2026 shutdown with Medium 3.1, so both generations of the id go dark on the same day rather than the older one leaving first.", "deprecated_on": "2026-05-22", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "mistral", "model": "mistral-medium-2508", "name": "Mistral Medium 3.1", "aliases": [], "description": "The August 2025 refresh of Mistral Medium 3, and the model most Mistral API traffic sat on through the second half of that year. It is deprecated but not yet gone: Mistral gave the 3.x Mediums an unusually long runway, announcing on 22 May 2026 and setting the shutdown for the end of August, so this is the one entry in the wave you can still call today.", "deprecated_on": "2026-05-22", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "note": "Medium 3.5 is multimodal where 3.1 was text-only, so a request that used to be rejected for containing an image part will now succeed and be billed.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "name": "Mistral Medium 3.5", "aliases": [], "description": "Mistral's frontier-class multimodal model, and the row every other row in the deprecation table points at. Large 2.1, Pixtral Large, all three Magistral Medium releases and both Devstral generations were retired into this one id, because Mistral decided reasoning, vision and coding did not each need their own endpoint. If you are migrating anything premier-tier off Mistral, this is almost certainly where the arrow lands.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "mistral-moderation-2411", "name": "Mistral Moderation", "aliases": [], "description": "The classifier behind Mistral's moderation endpoint, scoring text across nine policy categories. It shipped in November 2024 and ran nineteen months, which makes it one of the longest-lived ids Mistral has had - moderation models get refreshed on policy changes rather than on capability jumps.", "deprecated_on": "2026-03-31", "shutdown_on": "2026-06-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-moderation-26-03", "recommended": true, "note": "Check the category list before swapping. Moderation 2 scores a different taxonomy, so a threshold tuned against the 24.11 category keys will not map across one to one.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-moderation-26-03", "name": "Mistral Moderation 2", "aliases": [], "description": "The classifier behind Mistral's moderation endpoint, and the replacement for the 24.11 original that shut down on 30 June 2026. Moderation models are the ones people forget to migrate, because they usually sit in a middleware layer nobody has opened in a year - and a moderation call that starts erroring either fails open or blocks every request, depending on how you wrote it.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "mistral-ocr-2503", "name": "Mistral OCR", "aliases": [], "description": "The original Mistral OCR, launched in March 2025 and marketed as the best document-understanding API on the market at a thousand pages per dollar. It was gone by the end of that year. Its deprecation notice gave twenty-nine days, the shortest runway of any Mistral model with both dates published.", "deprecated_on": "2025-12-02", "shutdown_on": "2025-12-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-ocr-4-0", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-ocr-2505", "name": "Mistral OCR 2", "aliases": [], "description": "The second OCR release, adding better table and equation handling to the model that had made Mistral's document API worth using. It ran a year before the February 2026 wave took it. Both OCR ids that came before OCR 4 are now dead, which makes this the fastest-rotating family Mistral runs.", "deprecated_on": "2026-02-27", "shutdown_on": "2026-05-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-ocr-4-0", "recommended": true, "note": "OCR 4 returns structural block labels and paragraph-level bounding boxes that OCR 2 did not emit, so downstream code that indexed into the response by position rather than by key needs revisiting.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-ocr-4-0", "name": "Mistral OCR 4", "aliases": [], "description": "Mistral's document-understanding service, now returning paragraph-level bounding boxes and structural block labels rather than a flat text dump. It is the third OCR id in two years and both predecessors are shut down, so this is a line that rotates fast: if you pinned mistral-ocr-2503 or mistral-ocr-2505 in a pipeline, neither answers any more.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "mistral-saba-2502", "name": "Mistral Saba", "aliases": [], "description": "A 24B model trained for Arabic and South Indian languages, the only regional model Mistral has ever put on the API. It was deprecated four months after launch and Mistral pointed it at Mistral Small 4, a general model with no particular claim on Arabic - if Saba was doing something specific for you, there is no like-for-like successor.", "deprecated_on": "2025-06-10", "shutdown_on": "2025-09-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-small-2402", "name": "Mistral Small 1.0", "aliases": [], "description": "The original Small, launched February 2024 next to Mistral Large as the cheap half of the pair, and internally a tuned Mixtral 8x7B rather than a dense model. It was deprecated the same November and given seven months, outliving several models that shipped after it.", "deprecated_on": "2024-11-30", "shutdown_on": "2025-06-16", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-small-2409", "name": "Mistral Small 2.0", "aliases": [], "description": "A 22B model released in September 2024 under Mistral's research licence, sold as the midpoint between Nemo and Large. It is the odd size in the family history: every Small before and after it is either 12B or 24B, and nothing Mistral ships today is 22B.", "deprecated_on": "2025-11-06", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-small-2501", "name": "Mistral Small 3.0", "aliases": [], "description": "The 24B rewrite that reset the Small line in January 2025, dropping the layer count for latency and landing at roughly Llama 3.3 70B quality at a third of the size. It was the model that made \"small\" mean 24B rather than 22B at Mistral, and it lasted ten months.", "deprecated_on": "2025-11-06", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-small-2503", "name": "Mistral Small 3.1", "aliases": [], "description": "The first Small with vision, released March 2025 under Apache 2.0 with a 128k window, and for a while the best open model that fitted on a single RTX 4090. It was caught in the 6 November 2025 sweep that retired four generations of Mistral Small on one day, twenty-four days after the notice.", "deprecated_on": "2025-11-06", "shutdown_on": "2025-11-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-small-2506", "name": "Mistral Small 3.2", "aliases": [], "description": "The last of the Mistral Small 3 line, a June 2025 point release that mostly fixed instruction-following and repetition problems in 3.1 rather than adding anything. It outlived its three predecessors by eight months and stopped answering on 31 July 2026, which closed out the 3.x Small series entirely.", "deprecated_on": "2026-04-30", "shutdown_on": "2026-07-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "note": "Small 4 is a hybrid reasoning model. It can emit thinking content that Small 3.2 never produced, so parsers that assumed the whole response body was the answer need checking.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "mistral-small-4-0-26-03", "name": "Mistral Small 4", "aliases": [], "description": "The hybrid small model that folded instruct, reasoning and coding into a single weight class, which is why magistral-small and devstral-small stopped existing as separate ids after it shipped. Every Mistral Small that came before it - 1.0, 2.0, 3.0, 3.1 and 3.2 - has now been retired into this row, along with Saba, Mixtral and the two 8x-series mixtures.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "open-codestral-mamba", "name": "Codestral Mamba 7B", "aliases": [], "description": "A 7B code model built on the Mamba2 state-space architecture instead of a transformer, promising linear-time inference over long contexts. It is the only non-transformer Mistral ever served, and the only entry in the table whose deprecation and retirement fall on the same day - 6 June 2025, no notice period at all.", "deprecated_on": "2025-06-06", "shutdown_on": "2025-06-06", "status": "retired", "replacements": [ { "provider": "mistral", "model": "codestral-25-08", "recommended": true, "note": "Codestral 25.08 is a standard transformer. The linear-time long-context behaviour that was the whole point of the Mamba variant does not carry over.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "open-mistral-7b", "name": "Mistral 7B", "aliases": [], "description": "The model that made Mistral's reputation: an Apache 2.0 7B released by BitTorrent magnet link in September 2023 that beat Llama 2 13B and became the base for a generation of fine-tunes. The API id covered versions 0.2 and 0.3 and stopped answering on 30 March 2025, though the weights are permanent.", "deprecated_on": "2024-11-30", "shutdown_on": "2025-03-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "ministral-3-8b-25-12", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "open-mistral-nemo-2407", "name": "Mistral Nemo 12B", "aliases": [ "open-mistral-nemo" ], "description": "Built with NVIDIA, Apache 2.0, and the model that introduced the Tekken tokenizer Mistral still uses. It is the longest-lived open id in the catalogue - two years on the API, outlasting three full Mistral Small generations - and was only retired in the May 2026 wave.", "deprecated_on": "2026-05-22", "shutdown_on": "2026-07-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "ministral-3-8b-25-12", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "open-mixtral-8x22b", "name": "Mixtral 8x22B", "aliases": [], "description": "141B total parameters with 39B active, released April 2024 as the largest open model of its moment and the high-water mark for Mistral's fully-open strategy before Large 1.0 went closed a few weeks later. Retired 30 March 2025 alongside the rest of the open-prefix ids.", "deprecated_on": "2024-11-30", "shutdown_on": "2025-03-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "open-mixtral-8x7b", "name": "Mixtral 8x7B", "aliases": [], "description": "The first sparse mixture-of-experts anyone could download, dropped in December 2023 with no announcement, and the model that made MoE a mainstream open-weights architecture rather than a lab curiosity. Its API id was retired in March 2025; its influence on everything Mistral ships now was not.", "deprecated_on": "2024-11-30", "shutdown_on": "2025-03-30", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-small-4-0-26-03", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "pixtral-12b-2409", "name": "Pixtral 12B", "aliases": [], "description": "Mistral's first multimodal model, released Apache 2.0 in September 2024 and notable for handling images at native resolution rather than downsampling them to a fixed grid. Mistral points it at Ministral 3 14B, the only small model in the current range that takes images at all.", "deprecated_on": "2025-12-02", "shutdown_on": "2025-12-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "ministral-3-14b-25-12", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "pixtral-large-2411", "name": "Pixtral Large", "aliases": [], "description": "A 124B multimodal model built on Mistral Large 2.1 with a 1B vision encoder bolted on, and Mistral's answer to GPT-4o on charts and documents. It retired on the same day as the text model it was built from, and Mistral has not shipped a Pixtral since - vision is now a property of Medium 3.5 rather than a separate family.", "deprecated_on": "2026-02-27", "shutdown_on": "2026-05-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "mistral-medium-3-5-26-04", "recommended": true, "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "voxtral-mini-2507", "name": "Voxtral Mini", "aliases": [], "description": "The first Voxtral, which did double duty as both an audio-understanding model and a transcription endpoint - Mistral's docs listed it twice, once under each job. That is exactly the ambiguity the 26.02 split was meant to fix, and both of the doc rows pointing at this id retired together on 31 May 2026.", "deprecated_on": "2026-02-27", "shutdown_on": "2026-05-31", "status": "retired", "replacements": [ { "provider": "mistral", "model": "voxtral-mini-transcribe-26-02", "recommended": true, "note": "Mistral names the transcription model as the successor for both of this id's roles. If you were using it for audio question answering rather than transcription, that is a capability change and not a version bump.", "external": false } ], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "mistral", "model": "voxtral-mini-transcribe-26-02", "name": "Voxtral Mini Transcribe 2", "aliases": [], "description": "The transcription half of Voxtral, split out as its own id and pre-trained for the job rather than being a general audio model asked politely. It replaces voxtral-mini-2507, which served both the audio-understanding and the transcription use case under one name until Mistral decided the two wanted different models.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "voxtral-mini-transcribe-realtime-26-02", "name": "Voxtral Mini Transcribe Realtime", "aliases": [], "description": "The streaming sibling of Voxtral Mini Transcribe 2, tuned for live captioning rather than batch files. It is a separate id rather than a flag on the batch model, which means a live-transcription integration has its own lifecycle to track even though the two models share a version stamp today.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "mistral", "model": "voxtral-tts-26-03", "name": "Voxtral TTS", "aliases": [], "description": "Mistral's text-to-speech model, with zero-shot voice cloning and multilingual output. It is the newest branch of the Voxtral family and the first one that goes the other direction, generating audio rather than consuming it, so it has no predecessor in the deprecation table and nothing has been retired into it yet.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.mistral.ai/getting-started/models/models_overview/", "title": "Mistral models overview", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "moonshot", "model": "kimi-k2-0711-preview", "name": "Kimi K2 (0711)", "aliases": [], "description": "The first Kimi K2, a 1T-parameter mixture of experts released open-weights in July 2025 that briefly made Moonshot the strongest non-Western lab on agentic benchmarks. Its API id was discontinued on 25 May 2026 along with the rest of the K2 series; the weights remain downloadable.", "shutdown_on": "2026-05-25", "status": "retired", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "note": "Moonshot's guidance is to move to the latest Kimi model. K2.6 is the direct general-purpose descendant; K3 is the flagship.", "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "moonshot", "model": "kimi-k2-0905-preview", "name": "Kimi K2 (0905)", "aliases": [], "description": "The September 2025 K2 refresh, doubling the context window to 256k and sharply improving front-end code generation. It is the snapshot most K2-era agent frameworks pinned, and it went dark on 25 May 2026 with the rest of the series rather than being kept alive as the last good K2.", "shutdown_on": "2026-05-25", "status": "retired", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "moonshot", "model": "kimi-k2-thinking", "name": "Kimi K2 Thinking", "aliases": [], "description": "Moonshot's reasoning K2, able to run hundreds of sequential tool calls without losing the thread, and the model that put an open-weights system at the top of several agentic leaderboards in late 2025. Its id was retired on 25 May 2026 once thinking became a mode on the general models rather than a separate name.", "shutdown_on": "2026-05-25", "status": "retired", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "note": "K2.6 has a thinking mode you turn on per request. The response shape differs from this id's, so anything parsing reasoning content out of the body needs checking.", "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "moonshot", "model": "kimi-k2-thinking-turbo", "name": "Kimi K2 Thinking Turbo", "aliases": [], "description": "The fast-serving variant of K2 Thinking, and the fourth K2 id to express a combination of one model, one mode and one speed tier. Five ids for one model family is what the K2.6 consolidation was reacting to. Discontinued 25 May 2026.", "shutdown_on": "2026-05-25", "status": "retired", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "moonshot", "model": "kimi-k2-turbo-preview", "name": "Kimi K2 Turbo", "aliases": [], "description": "The high-throughput K2, same model served several times faster at a premium price, which made it the usual choice for interactive coding tools. Discontinued 25 May 2026. The pattern survives in kimi-k2.7-code-highspeed, so the id is gone but the tier is not.", "shutdown_on": "2026-05-25", "status": "retired", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.7-code-highspeed", "recommended": false, "note": "Moonshot points the whole K2 series at its latest models generically. If throughput was why you were on turbo, the code-highspeed id is the direct equivalent; if you wanted a general model, kimi-k2.6 is the closer fit.", "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "moonshot", "model": "kimi-k2.5", "name": "Kimi K2.5", "aliases": [], "description": "The general-intelligence model that carried Moonshot's API through the K2.x era. After the K3 launch it stopped being available to newly registered users while continuing to serve existing ones, and it goes dark with the rest of the legacy platform on 31 August. Moonshot's docs print that date without a year; read in context it is 2026.", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "note": "K2.6 is the closest general-purpose model and takes the same 256k of context. If you were relying on K2.5 being cheap rather than on what it could do, check pricing before switching - the newer ids are not like-for-like on cost.", "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "moonshot", "model": "kimi-k2.6", "name": "Kimi K2.6", "aliases": [], "description": "A 256k multimodal model taking text, image and video input, with thinking as a request-level mode rather than a separate id - the change that made kimi-k2-thinking and its turbo variant redundant. It is the general-purpose model in the current range, sitting below K3 and beside the code-specific K2.7s.", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "moonshot", "model": "kimi-k2.7-code", "name": "Kimi K2.7 Code", "aliases": [], "description": "The coding-specialised model of the K2.7 generation, 256k of context and tuned for agentic edit-and-run loops rather than single completions. Moonshot has kept a code-specific id in every generation since K2 even as it merged thinking and non-thinking modes into single models.", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "moonshot", "model": "kimi-k2.7-code-highspeed", "name": "Kimi K2.7 Code Highspeed", "aliases": [], "description": "The same K2.7 Code weights served on faster infrastructure at 180 to 260 tokens a second, sold as a separate id rather than a priority flag. Moonshot has used this turbo-suffix pattern since the K2 era, and it means a latency change is a model migration with its own lifecycle to track.", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "moonshot", "model": "kimi-k3", "name": "Kimi K3", "aliases": [], "description": "Moonshot's flagship, a 2.8T-parameter model with a 1M-token window, native visual understanding and a focus on long-horizon coding work. Its launch is what closed kimi-k2.5 and the entire moonshot-v1 series to new users, so it is both the current model and the direct cause of the largest retirement in Moonshot's history.", "status": "active", "replacements": [], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "moonshot", "model": "kimi-latest", "name": "Kimi Latest", "aliases": [], "description": "A floating alias that always resolved to whatever model the Kimi consumer app was running, with automatic context caching attached. Moonshot discontinued it on 28 January 2026, which is the unusual part: most providers keep a latest pointer forever precisely so they never have to retire one.", "shutdown_on": "2026-01-28", "status": "retired", "replacements": [ { "provider": "moonshot", "model": "kimi-k3", "recommended": true, "note": "There is no floating id in the current range. You now have to name a specific model, which means picking up new releases is a deliberate change rather than something that happens to you overnight.", "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "moonshot", "model": "kimi-thinking-preview", "name": "Kimi Thinking Preview", "aliases": [], "description": "Moonshot's first reasoning model, a multimodal preview released in May 2025 at a price several times the rest of the range. It is the oldest entry in Moonshot's discontinued table, switched off on 11 November 2025, six months after it appeared.", "shutdown_on": "2025-11-11", "status": "retired", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "moonshot", "model": "moonshot-v1-128k", "name": "Moonshot v1 128k", "aliases": [], "description": "The long-context member of Moonshot's original API family, and the model that built the company's reputation for handling very long documents back when 128k was a headline number. Context length was part of the model id rather than a runtime property, which is why there are three of these. Closed to new users after the K3 launch and sunsetting on 31 August, printed in Moonshot's docs without a year.", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "note": "Nothing in the current range encodes context length in the id. Pick the model, not the window - K2.6 gives you 256k and K3 gives you 1M.", "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "moonshot", "model": "moonshot-v1-128k-vision-preview", "name": "Moonshot v1 128k Vision Preview", "aliases": [], "description": "Moonshot's long-context vision model, still carrying a preview suffix at the point it was scheduled for shutdown - it never graduated to a stable id in more than a year of service. Closed to new users after K3 and sunsetting on 31 August with the rest of the v1 platform.", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "note": "K2.6 takes images and video as well as text, so the vision-preview split disappears entirely.", "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "moonshot", "model": "moonshot-v1-32k", "name": "Moonshot v1 32k", "aliases": [], "description": "The middle tier of the original Moonshot API, priced between the 8k and 128k variants of the same underlying model. Choosing between the three was a billing decision rather than a capability one, a design Moonshot abandoned entirely with the Kimi-branded ids. It sunsets on 31 August with the rest of the v1 line.", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "moonshot", "model": "moonshot-v1-32k-vision-preview", "name": "Moonshot v1 32k Vision Preview", "aliases": [], "description": "The 32k image-capable variant of the v1 line. Between three context sizes and two modalities the original API needed six ids to express what the current range does with one model, which is the whole argument for the K-series renaming. Sunsets 31 August.", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "moonshot", "model": "moonshot-v1-8k", "name": "Moonshot v1 8k", "aliases": [], "description": "The cheapest and oldest id Moonshot serves, the short-context entry point to the original API and the one most quickstart tutorials still reference. Two years of copy-pasted example code points at this string, which makes its 31 August sunset the most disruptive line in Moonshot's model list.", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "moonshot", "model": "moonshot-v1-8k-vision-preview", "name": "Moonshot v1 8k Vision Preview", "aliases": [], "description": "The short-context vision preview, and the tightest fit of the six v1 ids: 8k of context does not go far once an image is tokenised. It is closed to new registrations and shuts down on 31 August alongside every other moonshot-v1 string.", "shutdown_on": "2026-08-31", "status": "deprecated", "replacements": [ { "provider": "moonshot", "model": "kimi-k2.6", "recommended": true, "external": false } ], "sources": [ { "url": "https://platform.kimi.ai/docs/models", "title": "Kimi model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "openai", "model": "ada", "name": "ada", "aliases": [], "description": "The smallest and fastest GPT-3 model, priced so low it was used for embeddings and simple classification at a scale nothing else made sense for. Retired January 2024; the embeddings side of its job moved to the dedicated text-embedding models.", "released_on": "2020-06-11", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "babbage-002", "recommended": true, "note": "For embeddings specifically, text-embedding-3-small is the modern equivalent and is not deprecated.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "babbage", "name": "babbage", "aliases": [], "description": "The second-smallest GPT-3 base model, mostly used for semantic search and cheap classification where latency mattered more than nuance. Retired January 2024 with the rest of the original GPT-3 line.", "released_on": "2020-06-11", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "babbage-002", "recommended": true, "note": "babbage-002 was the direct replacement and is deprecated in turn, shutting down September 2026.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "babbage-002", "name": "babbage-002", "aliases": [], "description": "A base completions model kept in service almost entirely as a fine-tuning target for classification work, where a small model trained on your own labels still beat prompting a large one. Its September 2026 shutdown ends that option.", "released_on": "2023-08-22", "deprecated_on": "2025-09-26", "shutdown_on": "2026-09-28", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.4-mini", "recommended": true, "note": "Existing fine-tunes do not transfer. Budget for re-running training on the new base and re-validating accuracy — a base-model fine-tune and an instruction-model fine-tune do not behave the same way.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "chatgpt-4o-latest", "name": "ChatGPT-4o latest", "aliases": [], "description": "The id that exposed the exact GPT-4o variant serving ChatGPT, complete with the personality tuning the API models did not have. It was popular for chat products that wanted ChatGPT's voice, and it was retired in February 2026.", "deprecated_on": "2025-11-18", "shutdown_on": "2026-02-17", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.1-chat-latest", "recommended": true, "note": "OpenAI named gpt-5.1-chat-latest as the successor — but that id has since been retired too. The chain currently ends at gpt-5.6-sol.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "chatgpt-image-latest", "name": "ChatGPT Image latest", "aliases": [], "description": "A floating pointer to whichever image model ChatGPT was using, for products that wanted to match what users saw in the app. Like every other -latest pointer OpenAI has shipped, it was retired on a shorter notice than the pinned ids it pointed at.", "deprecated_on": "2026-06-02", "shutdown_on": "2026-12-01", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-image-2", "recommended": true, "note": "Move to a pinned id. Every floating pointer in this catalog — chat, codex and image alike — has now been deprecated at least once.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "code-davinci-002", "name": "code-davinci-002", "aliases": [], "description": "The Codex model behind the original GitHub Copilot, and the shortest notice OpenAI has ever given: deprecated on March 20, 2023 and shut down three days later. The backlash from researchers who lost access mid-study is a large part of why every provider now publishes deprecation schedules months ahead.", "deprecated_on": "2023-03-20", "shutdown_on": "2023-03-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-4o", "recommended": true, "note": "For code work today, gpt-5.3-codex is the direct descendant of this line. gpt-4o is what OpenAI named at the time.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "codex-mini-latest", "name": "codex-mini-latest", "aliases": [], "description": "The small coding model behind the first Codex CLI, and a floating pointer rather than a pinned snapshot. Retired February 2026, a day short of three months after the announcement.", "deprecated_on": "2025-11-17", "shutdown_on": "2026-02-12", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5-codex-mini", "recommended": true, "note": "gpt-5-codex-mini keeps the shell-tool integration this model was built around, so agent scaffolding usually transfers unchanged.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "computer-use-preview", "name": "computer-use-preview", "aliases": [ "computer-use-preview-2025-03-11" ], "description": "The model behind OpenAI's Operator agent, which drove a browser by emitting mouse and keyboard actions against screenshots. It stayed a preview for its entire life and was retired in July 2026 without a direct replacement.", "released_on": "2025-03-11", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-terra", "recommended": true, "note": "There is no drop-in computer-use id in the current generation. The computer_use tool schema does not carry over — this is a rewrite, not a swap.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "curie", "name": "curie", "aliases": [], "description": "The mid-size GPT-3 base model, roughly a tenth of davinci's price and the usual choice for classification once you had few-shot examples that worked. Its January 2024 shutdown removed the middle of the GPT-3 size ladder.", "released_on": "2020-06-11", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "davinci-002", "recommended": true, "note": "OpenAI collapsed four sizes into two. There is no curie-sized base model on the platform now.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "dall-e-2", "name": "DALL·E 2", "aliases": [], "description": "The model that put text-to-image generation in front of a general audience in 2022, complete with the waitlist that made it feel like an event. By the time it was retired in May 2026 it had been outclassed for years — it survived on inertia and on the inpainting API that had no direct successor for a while.", "released_on": "2022-04-06", "deprecated_on": "2025-11-14", "shutdown_on": "2026-05-12", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-image-2", "recommended": true, "note": "The request shape differs: gpt-image models are called through the images endpoint with a different parameter set, and there is no direct equivalent of the DALL·E 2 variations call.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "dall-e-3", "name": "DALL·E 3", "aliases": [], "description": "The image model that finally rendered text inside images legibly and took prompts as prose rather than keyword soup. It rewrote your prompt before generating, which produced better pictures and made exact reproduction impossible — a trade the gpt-image line kept.", "released_on": "2023-10-03", "deprecated_on": "2025-11-14", "shutdown_on": "2026-05-12", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-image-2", "recommended": true, "note": "Quality is a clear step up, but the automatic prompt rewriting behaves differently. Anything with prompts tuned against DALL·E 3s rewriter needs re-testing rather than porting.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "davinci", "name": "davinci", "aliases": [], "description": "The largest GPT-3 base model, and the one the 2020 paper actually described. It had no instruction tuning at all — you steered it with few-shot examples, which is the technique the whole prompt-engineering field grew out of. Retired January 2024.", "released_on": "2020-06-11", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "davinci-002", "recommended": true, "note": "davinci-002 was the named replacement base model and is itself deprecated, shutting down September 2026 — after which OpenAI hosts no base models at all.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "davinci-002", "name": "davinci-002", "aliases": [], "description": "The larger of the two remaining base completions models, and the last descendant of the original davinci line still callable. Deprecated with a September 2026 shutdown.", "released_on": "2023-08-22", "deprecated_on": "2025-09-26", "shutdown_on": "2026-09-28", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.4-mini", "recommended": true, "note": "Nothing on the current platform offers raw base-model continuation. If your use case depends on it rather than on instruction following, an open-weight model is a closer match than anything OpenAI still hosts.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-3.5-turbo", "name": "GPT-3.5 Turbo", "aliases": [ "gpt-3.5-turbo-0125", "gpt-3.5-turbo-completions" ], "description": "The model that made the chat completions API a commodity: cheap enough to put behind a free product, good enough that most people stopped writing prompt chains. It survived every deprecation wave from 2023 to 2026 because so much code was written against it — and it is finally going in October 2026.", "released_on": "2023-03-01", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-terra", "recommended": true, "note": "The biggest gotcha is max_tokens: current models use max_completion_tokens and reject the old parameter. Prompts tuned for GPT-3.5 are usually far more verbose than a 5.6-generation model needs.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-3.5-turbo-0301", "name": "GPT-3.5 Turbo (0301)", "aliases": [], "description": "The launch snapshot of GPT-3.5 Turbo, from March 2023 — the model that made chat APIs cheap and kicked off the current wave of applications. It was deprecated three months after release and took eighteen months to actually go.", "released_on": "2023-03-01", "deprecated_on": "2023-06-13", "shutdown_on": "2024-09-13", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-3.5-turbo", "recommended": true, "note": "Now deprecated in turn. Anything still pinned to a 2023 snapshot should go straight to a current model rather than to the pointer.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-3.5-turbo-0613", "name": "GPT-3.5 Turbo (0613)", "aliases": [ "gpt-3.5-turbo-16k-0613" ], "description": "The June 2023 snapshot that introduced function calling to the API — the feature every agent framework since has been built on. Retired September 2024 along with its 16k variant.", "released_on": "2023-06-13", "deprecated_on": "2023-11-06", "shutdown_on": "2024-09-13", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-3.5-turbo", "recommended": true, "note": "The undated pointer was the named successor and is itself now deprecated, shutting down October 2026. Plan for gpt-5.6-terra instead.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-3.5-turbo-1106", "name": "GPT-3.5 Turbo (1106)", "aliases": [], "description": "The November 2023 GPT-3.5 snapshot, the first to support JSON mode and parallel function calling. It is deprecated with a September 2026 shutdown, a month before the undated gpt-3.5-turbo pointer follows it.", "released_on": "2023-11-06", "deprecated_on": "2025-09-26", "shutdown_on": "2026-09-28", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.4-mini", "recommended": true, "note": "Swap max_tokens for max_completion_tokens, and replace JSON mode with structured outputs, which enforce your schema rather than only valid JSON.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-3.5-turbo-instruct", "name": "GPT-3.5 Turbo Instruct", "aliases": [], "description": "The last completions-endpoint model OpenAI shipped, kept alive years past the rest of the /v1/completions family because logprobs and raw continuation behaviour had no equivalent in chat. Its September 2026 shutdown closes out the completions era entirely.", "released_on": "2023-09-18", "deprecated_on": "2025-09-26", "shutdown_on": "2026-09-28", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.4-mini", "recommended": true, "note": "This is an endpoint change, not just an id change: /v1/completions gives way to chat completions, prompts become messages, and raw continuation behaviour has no direct equivalent.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-4", "name": "GPT-4", "aliases": [ "gpt-4-0613", "gpt-4-completions", "gpt-4-0613-completions" ], "description": "The model that started the current era, and still the id people type when they mean \"the good one\". OpenAI kept it alive for three and a half years — far longer than anything released since — and set its shutdown for October 2026.", "released_on": "2023-03-14", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Sol is a reasoning model, so it is not a behavioural drop-in: swap max_tokens for max_completion_tokens and expect reasoning tokens on the bill. Prompts written to coax GPT-4 into thinking step by step are now counterproductive.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-4-0125-preview", "name": "GPT-4 Turbo preview (0125)", "aliases": [], "description": "The January 2024 preview snapshot, released to fix a \"laziness\" regression where the model truncated long code generations. Retired in March 2026.", "released_on": "2024-01-25", "deprecated_on": "2025-09-26", "shutdown_on": "2026-03-26", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Nothing about the preview snapshots carries forward; the response-truncation behaviour they existed to fix has not recurred.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4-0314", "name": "GPT-4 (0314)", "aliases": [], "description": "The launch snapshot of GPT-4, from March 2023. It was scheduled for removal twice — first for June 2024, then extended — before finally going in March 2026, which made it one of the longest-lived pinned snapshots OpenAI ever shipped.", "released_on": "2023-03-14", "deprecated_on": "2025-09-26", "shutdown_on": "2026-03-26", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-4", "recommended": true, "note": "The undated gpt-4 id is the smallest change, though it is itself deprecated with an October 2026 shutdown. For a migration that lasts, go to gpt-5.6-sol.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4-1106-preview", "name": "GPT-4 Turbo preview (1106)", "aliases": [], "description": "The November 2023 preview that introduced 128k context and JSON mode, and the one entry on this site with two conflicting shutdown dates: OpenAI first set it for March 26, 2026, then listed the same id again in the April 2026 announcement with a new date of October 23, 2026. The later announcement is the one that stands.", "released_on": "2023-11-06", "deprecated_on": "2025-09-26", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "JSON mode has been superseded by structured outputs, which enforce a schema rather than only guaranteeing valid JSON. Worth adopting during the move instead of after it.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-4-32k", "name": "GPT-4 32k", "aliases": [ "gpt-4-32k-0613", "gpt-4-32k-0314" ], "description": "The 32,000-token variant of the original GPT-4, sold at twice the price of the 8k model when a long context was still a premium feature. Access was gated behind a waitlist for most of its life, so it never had the reach of plain gpt-4. GPT-4o made it redundant: 128k of context, faster, and cheaper than the 8k model it replaced.", "released_on": "2023-06-13", "deprecated_on": "2024-06-06", "shutdown_on": "2025-06-06", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-4o", "recommended": true, "note": "GPT-4o takes 128k of context, so nothing that fit in 32k needs chunking. It is a Chat Completions model with the same message format — the only required change is the id.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4-turbo", "name": "GPT-4 Turbo", "aliases": [ "gpt-4-turbo-2024-04-09", "gpt-4-turbo-completions" ], "description": "The 128k-context, cheaper GPT-4 that made long documents practical and held the default slot until GPT-4o arrived a month after its final snapshot. It is deprecated with an October 2026 shutdown.", "released_on": "2024-04-09", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Context is no longer the constraint it was, so chunking logic written for 128k can usually be deleted rather than ported.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-4-turbo-preview", "name": "GPT-4 Turbo preview", "aliases": [ "gpt-4-turbo-preview-completions" ], "description": "A floating pointer that resolved to whichever GPT-4 Turbo preview snapshot was current, which meant its behaviour changed underneath anyone who used it. It was retired in March 2026 along with the snapshots it pointed at.", "deprecated_on": "2025-09-26", "shutdown_on": "2026-03-26", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Prefer a pinned id this time. Floating preview pointers have been retired in every OpenAI deprecation wave since 2024.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4-vision-preview", "name": "GPT-4 Vision preview", "aliases": [ "gpt-4-1106-vision-preview" ], "description": "The first GPT-4 that could see, and for a few months the only way to get image understanding from OpenAI. It was a separate id with its own rate limits and no function calling — constraints that disappeared when GPT-4o folded vision into the main model. Retired December 2024.", "released_on": "2023-11-06", "deprecated_on": "2024-06-06", "shutdown_on": "2024-12-06", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-4o", "recommended": true, "note": "Vision is no longer a separate id. The image_url content block carries over unchanged, and function calling now works alongside it.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4.1", "name": "GPT-4.1", "aliases": [], "description": "The 2025 refresh that fixed what GPT-4o was weakest at — instruction following and long-context recall — and became the default non-reasoning model for a year. Its nano sibling has been deprecated; gpt-4.1 and gpt-4.1-mini have not.", "released_on": "2025-04-14", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-4.1-mini", "name": "GPT-4.1 mini", "aliases": [], "description": "The mid-size GPT-4.1 variant, and one of the few 2025-era ids that has survived the 2026 deprecation waves untouched. It is not listed on the OpenAI deprecations page in any announcement, unlike gpt-4.1-nano.", "released_on": "2025-04-14", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-4.1-nano", "name": "GPT-4.1 nano", "aliases": [ "gpt-4.1-nano-2025-04-14" ], "description": "The cheapest model of the GPT-4.1 generation, and the only member of that family to be deprecated so far — gpt-4.1 and gpt-4.1-mini remain active. It was a common choice for classification and routing at high volume.", "released_on": "2025-04-14", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-luna", "recommended": true, "note": "Fine-tuned gpt-4.1-nano deployments are directed to gpt-5.4-nano instead; a fine-tune does not transfer and has to be re-run on the new base.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-4.5-preview", "name": "GPT-4.5 preview", "aliases": [], "description": "The largest non-reasoning model OpenAI ever shipped, released in February 2025 and withdrawn from the API three months later — the shortest-lived flagship in the platform's history. It was expensive to serve and was overtaken by reasoning models that got more from less compute.", "released_on": "2025-02-27", "deprecated_on": "2025-04-14", "shutdown_on": "2025-07-14", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-4.1", "recommended": true, "note": "gpt-4.1 was the named successor and is still active — one of the few migration targets from this era that has not itself been deprecated.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o", "name": "GPT-4o", "aliases": [], "description": "The omni model that made multimodal input routine and halved the price of GPT-4 Turbo at the same time. It is still the default many production stacks reach for, and OpenAI has kept the floating gpt-4o pointer alive even as the dated snapshot behind it changed. The May 2024 snapshot is deprecated; the pointer itself is not.", "released_on": "2024-05-13", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-4o-2024-05-13", "name": "GPT-4o (2024-05-13)", "aliases": [], "description": "The original GPT-4o snapshot, the one that shipped at the May 2024 launch. Only this dated snapshot is deprecated — the floating gpt-4o pointer is not, and now resolves to a later snapshot. If you pinned the date, this is your deadline; if you did not, you have already been migrated.", "released_on": "2024-05-13", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "If the pin was defensive rather than deliberate, moving to the floating gpt-4o id is the smaller change and buys time.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-4o-audio-preview", "name": "GPT-4o Audio preview", "aliases": [], "description": "Audio in and audio out over ordinary Chat Completions, without the WebSocket session the realtime models require. It suited anything that did not need to be interrupted mid-sentence — voice notes, summaries, asynchronous replies.", "deprecated_on": "2025-09-15", "shutdown_on": "2026-05-07", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-audio-1.5", "recommended": true, "note": "Same request/response shape, so this is closer to a drop-in than the realtime migration. Re-check the supported input and output audio formats.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o-audio-preview-2024-10-01", "name": "GPT-4o Audio preview (2024-10-01)", "aliases": [], "description": "The first audio-capable Chat Completions snapshot, shipped alongside the original realtime preview. It was retired in October 2025, a year to the day after release.", "released_on": "2024-10-01", "deprecated_on": "2025-06-10", "shutdown_on": "2025-10-10", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-audio-1.5", "recommended": true, "note": "The audio content-block format changed after this snapshot; check how you encode and read back audio parts.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o-mini", "name": "GPT-4o mini", "aliases": [], "description": "The model that made GPT-4-class quality cheap enough to put in a loop, and for a long stretch the highest-volume id on the OpenAI platform. The floating pointer is not deprecated; only the dated gpt-4o-2024-05-13 snapshot of its larger sibling is.", "released_on": "2024-07-18", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-4o-mini-audio-preview", "name": "GPT-4o mini Audio preview", "aliases": [], "description": "The cost tier for audio on Chat Completions, aimed at transcription-adjacent work at volume. Retired in May 2026 with the rest of the gpt-4o audio family.", "deprecated_on": "2025-09-15", "shutdown_on": "2026-05-07", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-audio-mini", "recommended": true, "note": "gpt-audio-mini is deprecated in turn, shutting down January 2027. If you are migrating now, gpt-audio-1.5 avoids doing this twice.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o-mini-realtime-preview", "name": "GPT-4o mini Realtime preview", "aliases": [], "description": "The cheaper realtime preview, which made always-on voice affordable enough for consumer products — realtime audio is billed per minute, so the tier below the flagship mattered more here than anywhere else.", "deprecated_on": "2025-09-15", "shutdown_on": "2026-05-07", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-realtime-mini", "recommended": true, "note": "Note that gpt-realtime-mini is itself deprecated with a January 2027 shutdown; gpt-realtime-2.1-mini is the target that lasts.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o-mini-search-preview-2025-03-11", "name": "GPT-4o mini search-preview", "aliases": [], "description": "The cost tier of the search-preview pair, for grounded answers at volume. Retired on the same schedule as its larger sibling once web search became a tool rather than a model variant.", "released_on": "2025-03-11", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-terra", "recommended": true, "note": "Use the web_search tool on a current model. Search calls are billed separately from tokens, so re-model the unit cost before rolling out.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o-mini-transcribe-2025-03-20", "name": "GPT-4o mini Transcribe (2025-03-20)", "aliases": [], "description": "The March 2025 snapshot of OpenAI's small speech-to-text model, and the first thing to beat Whisper on word error rate on OpenAI's own API. It is the only transcription id in the July 2026 legacy-audio deprecation: the floating gpt-4o-mini-transcribe slug moved to the December snapshot, and this pinned one is what stops answering.", "deprecated_on": "2026-07-20", "shutdown_on": "2027-01-20", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-4o-mini-transcribe-2025-12-15", "recommended": true, "note": "If your code says gpt-4o-mini-transcribe without a date you are already on the December snapshot and nothing breaks. Only the pinned March string is going away.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI deprecations", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-4o-mini-transcribe-2025-12-15", "name": "GPT-4o mini Transcribe (2025-12-15)", "aliases": [ "gpt-4o-mini-transcribe" ], "description": "The December 2025 transcription snapshot, and what the undated gpt-4o-mini-transcribe slug resolves to today. OpenAI recommends it over the full-size gpt-4o-transcribe, which is unusual - the mini model is the default advice rather than the budget option.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI deprecations", "accessed": "2026-08-03" }, { "url": "https://developers.openai.com/api/docs/models", "title": "OpenAI models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "openai", "model": "gpt-4o-realtime-preview", "name": "GPT-4o Realtime preview", "aliases": [ "gpt-4o-realtime-preview-2025-06-03", "gpt-4o-realtime-preview-2024-12-17" ], "description": "The model that introduced speech-to-speech over a WebSocket, with interruption handling and sub-second response — the first time a voice agent could be built without stitching together transcription, a chat model and text-to-speech. It stayed a preview for its entire eighteen-month life.", "released_on": "2024-10-01", "deprecated_on": "2025-09-15", "shutdown_on": "2026-05-07", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-realtime-1.5", "recommended": true, "note": "The session event schema changed between preview and GA. Expect to update event handlers, not just the model id.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o-realtime-preview-2024-10-01", "name": "GPT-4o Realtime preview (2024-10-01)", "aliases": [], "description": "The original realtime snapshot from the October 2024 launch, deprecated eight months later on a four-month notice. Realtime previews were versioned by date and turned over faster than anything else on the platform at the time.", "released_on": "2024-10-01", "deprecated_on": "2025-06-10", "shutdown_on": "2025-10-10", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-realtime-1.5", "recommended": true, "note": "Two protocol revisions separate this snapshot from the current realtime model; treat the voice pipeline as a rewrite.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o-search-preview-2025-03-11", "name": "GPT-4o search-preview", "aliases": [], "description": "A GPT-4o variant with web search wired in at the model level, so a single call returned an answer with citations. OpenAI later moved search to a tool available on every model, which made the dedicated id redundant.", "released_on": "2025-03-11", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-terra", "recommended": true, "note": "Enable the web_search tool on any current model. The response shape differs: citations arrive as tool output rather than inline annotations.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-4o-transcribe", "name": "GPT-4o Transcribe", "aliases": [], "description": "The full-size speech-to-text model of the GPT-4o generation, still served and still undated. OpenAI's own guidance points at the mini model instead for best results, which leaves this id in the odd position of being the larger, more expensive option that the docs do not recommend.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/models", "title": "OpenAI models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5", "name": "GPT-5", "aliases": [ "gpt-5-2025-08-07" ], "description": "The August 2025 launch that unified the GPT and o-series lines into one model with a built-in router, and the id an enormous amount of production code was written against. OpenAI deprecated the snapshot behind it in June 2026 with a six-month runway. If your code says \"gpt-5\" with no date, this is the entry that applies to you.", "released_on": "2025-08-07", "deprecated_on": "2026-06-11", "shutdown_on": "2026-12-11", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Sol is a reasoning model, so budget output with max_completion_tokens rather than max_tokens, and expect reasoning_effort to dominate both cost and latency. Re-tune it before comparing prices.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-5-chat-latest", "name": "GPT-5 chat-latest", "aliases": [], "description": "The first of the chat-latest pointers, tracking whatever non-reasoning model ChatGPT was serving. It was retired in July 2026 with three months notice, along with the entire first-generation Codex line.", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Sol is pinned rather than floating, so behaviour stops changing underneath you between deploys.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-5-codex", "name": "GPT-5 Codex", "aliases": [], "description": "The model that relaunched the Codex name for agentic coding, tuned for long autonomous runs rather than inline completion. It was retired in July 2026 together with every other first-generation Codex id in a single announcement.", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "For coding-agent work the current equivalent is gpt-5.3-codex; OpenAI names Sol as the general replacement.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-5-codex-mini", "name": "GPT-5 Codex mini", "aliases": [], "description": "The small coding model that replaced codex-mini-latest when that pointer was retired in February 2026. It kept the shell-tool integration the Codex line was built around, at a size that suits editor-loop completions rather than long-running agent work.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5-mini", "name": "GPT-5 mini", "aliases": [ "gpt-5-mini-2025-08-07" ], "description": "The cost tier of the original GPT-5 release, and the model that absorbed most of the volume GPT-4o mini used to carry. It was deprecated in the same June 2026 announcement as the rest of the GPT-5 family, on the same December shutdown date.", "released_on": "2025-08-07", "deprecated_on": "2026-06-11", "shutdown_on": "2026-12-11", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-terra", "recommended": true, "note": "Terra is the volume tier of the current generation, so this is the like-for-like swap. Check per-token pricing before assuming the migration is cost-neutral in either direction.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-5-nano", "name": "GPT-5 nano", "aliases": [ "gpt-5-nano-2025-08-07" ], "description": "The smallest model of the GPT-5 generation, priced for classification and routing work that runs millions of times a day. Deprecated alongside gpt-5 and gpt-5-mini in June 2026.", "released_on": "2025-08-07", "deprecated_on": "2026-06-11", "shutdown_on": "2026-12-11", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-luna", "recommended": true, "note": "Luna occupies the same slot. At nano volumes even a small change in output length moves the bill, so re-measure token counts on real traffic rather than trusting the per-token price alone.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-5-pro", "name": "GPT-5 Pro", "aliases": [ "gpt-5-pro-2025-10-06" ], "description": "The extended-compute tier of GPT-5, which spent far longer per request to buy accuracy on hard problems. It shipped two months after the base model and was deprecated on the same date, which is the useful lesson: a Pro tier does not get a longer life than the generation it belongs to.", "released_on": "2025-10-06", "deprecated_on": "2026-06-11", "shutdown_on": "2026-12-11", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Sol at a high reasoning_effort is the closest equivalent. There is no separate \"-pro\" id in the 5.6 generation — the compute dial moved into the request instead of the model name.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-5.1-chat-latest", "name": "GPT-5.1 chat-latest", "aliases": [], "description": "The pointer that briefly replaced chatgpt-4o-latest when that id was retired in February 2026 — and was itself retired five months later. Anyone who followed OpenAI's own migration advice in late 2025 landed here and had to move again almost immediately.", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Move to a pinned model rather than another chat-latest pointer; the pointer ids have now been retired three times in under a year.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-5.1-codex", "name": "GPT-5.1 Codex", "aliases": [], "description": "The second-generation Codex model, and the one most coding agents shipped against through late 2025. Retired July 23, 2026 alongside gpt-5-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini and gpt-5.2-codex — four generations of the same line, gone on one day.", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.3-codex", "recommended": true, "note": "gpt-5.3-codex is the current hosted Codex model. Given how fast this line turns over, pin it and set a calendar reminder rather than assuming a long life.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-5.1-codex-max", "name": "GPT-5.1 Codex Max", "aliases": [], "description": "The largest Codex variant, built for repository-scale work where the model needed to hold an entire codebase in context. It lasted about eight months before being retired with the rest of the Codex line.", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.3-codex", "recommended": true, "note": "There is no separate -max id in the 5.3 Codex generation; the capability moved into the base model.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-5.1-codex-mini", "name": "GPT-5.1 Codex mini", "aliases": [], "description": "The small Codex model for editor-loop completions, where a 200ms difference in latency is more valuable than a smarter answer. Retired July 23, 2026.", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-terra", "recommended": true, "note": "OpenAI names Terra rather than a Codex id here. For editor-latency work gpt-5.3-codex-spark is the closer fit.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-5.2-chat-latest", "name": "GPT-5.2 chat-latest", "aliases": [], "description": "One of the non-reasoning chat pointers that mirrored the model behind ChatGPT at a given moment. These ids exist for conversational parity, not for tool use, and OpenAI retires them faster than the numbered API models — this one got a three-month notice, not six.", "deprecated_on": "2026-05-08", "shutdown_on": "2026-08-10", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "The chat-latest pointers do not support the full tool-calling surface, so moving to Sol is usually an upgrade rather than a like-for-like swap. Check any prompt that was tuned against ChatGPT-flavoured output.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-5.2-codex", "name": "GPT-5.2 Codex", "aliases": [], "description": "The third-generation Codex model, and the shortest-lived of the family: it was deprecated within months of release and retired on the same day as its three predecessors.", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.3-codex", "recommended": true, "note": "The direct successor. The Codex line has averaged well under a year per id — treat any Codex model as a moving target.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-5.3-chat-latest", "name": "GPT-5.3 chat-latest", "aliases": [], "description": "The successor pointer to gpt-5.2-chat-latest, deprecated on the same day and the same shutdown date as the id it replaced. Chat-latest pointers move with ChatGPT rather than with the API, which is exactly why pinning one is risky: the behaviour underneath can change without the id changing.", "deprecated_on": "2026-05-08", "shutdown_on": "2026-08-10", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Sol is a pinned model rather than a moving pointer, so output stays stable between deploys. That is the main behavioural difference to plan for.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-5.3-codex", "name": "GPT-5.3 Codex", "aliases": [], "description": "The current hosted Codex model, after gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max and gpt-5.2-codex were all retired together in July 2026. The Codex line turns over faster than any other family on the platform, which is the single most useful thing to know when pinning one of these ids.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.3-codex-spark", "name": "GPT-5.3 Codex Spark", "aliases": [], "description": "The low-latency Codex variant, built for the completion loop inside an editor rather than for long agent runs. It shares the Codex family's short release cadence: four of its predecessors were retired on a single day in July 2026.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.4", "name": "GPT-5.4", "aliases": [], "description": "For several months the default recommendation for professional work on the OpenAI platform, and still widely deployed. It is not deprecated, but two generations now sit above it, and OpenAI has been retiring GPT-5-era snapshots on roughly six-month notice.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.4-mini", "name": "GPT-5.4 mini", "aliases": [], "description": "The cost tier of the GPT-5.4 generation. It is the named replacement for the legacy fine-tuning base models — ft-gpt-3.5-turbo, ft-babbage-002 and ft-davinci-002 — which is an unusual role for a mini model and worth checking before assuming a fine-tune migrates cleanly.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.4-nano", "name": "GPT-5.4 nano", "aliases": [], "description": "The smallest model of the GPT-5.4 generation, and the successor OpenAI names for fine-tuned gpt-4.1-nano deployments. Nano-class models are where cost pressure is sharpest, which is also why they turn over faster than the tiers above them.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.4-pro", "name": "GPT-5.4 Pro", "aliases": [], "description": "The extended-compute variant of GPT-5.4, positioned for problems worth minutes rather than seconds. OpenAI retired the earlier gpt-5-pro snapshot on the same announcement as the rest of the GPT-5 family, so a Pro tier is no guarantee of a longer life than the base model.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.5", "name": "GPT-5.5", "aliases": [], "description": "The generation between GPT-5.4 and the Sol/Terra/Luna split, and the last one to use a plain numeric name. It has not been deprecated, but it sits directly behind models that have — worth knowing if a migration plan was written against it earlier in 2026.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.5-pro", "name": "GPT-5.5 Pro", "aliases": [], "description": "The extended-compute variant of GPT-5.5, which spends much longer per request in exchange for accuracy on problems where a wrong answer is expensive. Pro tiers have historically been deprecated on the same schedule as the base model they derive from, so its lifecycle is worth watching alongside GPT-5.5.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.6-luna", "name": "GPT-5.6 Luna", "aliases": [], "description": "The smallest and cheapest model of the GPT-5.6 generation, the slot GPT-4.1 nano used to occupy. OpenAI names it as the replacement for gpt-4.1-nano, so it inherits the classification, extraction and routing workloads that never needed a frontier model in the first place.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.6-sol", "name": "GPT-5.6 Sol", "aliases": [], "description": "The flagship of the Sol/Terra/Luna split, where OpenAI stopped shipping one frontier model in three sizes and started naming them for what they are for. Sol is the reasoning tier, and it is where almost every 2026 deprecation points: the o-series, the GPT-5 snapshots, the codex line and the chat-latest pointers all name it as their successor.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-5.6-terra", "name": "GPT-5.6 Terra", "aliases": [], "description": "The middle tier of the GPT-5.6 generation, aimed at high-volume work where latency and price matter more than the last few points of reasoning. It is the named replacement for o4-mini, the search-preview models and the long tail of GPT-3.5 ids, which makes it the busiest migration target on the OpenAI side of this catalog.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-audio", "name": "GPT Audio", "aliases": [], "description": "The GA audio model for Chat Completions, successor to the gpt-4o audio previews. Deprecated in July 2026 with a six-month runway, alongside the whole realtime and audio generation it belonged to.", "deprecated_on": "2026-07-20", "shutdown_on": "2027-01-20", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-audio-1.5", "recommended": true, "note": "A point upgrade within the same family — the closest thing to a drop-in in the audio line.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-audio-1.5", "name": "GPT Audio 1.5", "aliases": [], "description": "The current audio-in/audio-out model on Chat Completions, and the target OpenAI names for every deprecated audio id — the gpt-4o audio previews, gpt-audio and gpt-audio-mini all point here. Unlike the realtime models it is a normal request/response call.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-audio-mini", "name": "GPT Audio mini", "aliases": [], "description": "The cheap audio model, and the named replacement for gpt-4o-mini-audio-preview before being deprecated itself nine months later. Anyone following OpenAIs own migration advice through 2026 has now been moved twice.", "deprecated_on": "2026-07-20", "shutdown_on": "2027-01-20", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-audio-1.5", "recommended": true, "note": "There is no mini id in the 1.5 audio generation; the cheaper tier folded back into the base model.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-audio-mini-2025-10-06", "name": "GPT Audio mini (2025-10-06)", "aliases": [], "description": "The pinned October 2025 snapshot behind gpt-audio-mini, retired in July 2026 while the floating id it backed stayed alive until 2027. The dated snapshot going first is the opposite of what most people assume pinning buys them.", "released_on": "2025-10-06", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-audio-1.5", "recommended": true, "note": "Go straight to gpt-audio-1.5 rather than to the floating gpt-audio-mini, which shuts down in January 2027.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "gpt-image-1", "name": "GPT Image 1", "aliases": [], "description": "The first image model of the GPT era, and the one that made image editing with a reference picture routine. It replaced the DALL·E line and was itself deprecated less than a year later, in the same April 2026 announcement that took out the o-series.", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-image-2", "recommended": true, "note": "A same-family upgrade, so the request shape carries over. Check output pricing per image size before rolling out — the tiers changed.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-image-1-mini", "name": "GPT Image 1 mini", "aliases": [], "description": "The cost tier of the first GPT image generation, for thumbnails and drafts where a cheaper picture is the right answer. Deprecated in June 2026 with a December shutdown, two months after its larger sibling goes.", "deprecated_on": "2026-06-02", "shutdown_on": "2026-12-01", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-image-2", "recommended": true, "note": "There is no mini id in the gpt-image-2 generation. Cost control moves to the size and quality parameters instead of a separate model.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-image-1.5", "name": "GPT Image 1.5", "aliases": [], "description": "A late refresh of the first GPT image generation, deprecated barely after it landed — it and gpt-image-2 were announced close enough together that many teams never migrated onto it at all.", "deprecated_on": "2026-06-02", "shutdown_on": "2026-12-01", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-image-2", "recommended": true, "note": "If you are still on gpt-image-1, skip 1.5 entirely and go straight to gpt-image-2 — 1.5 shuts down first.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-image-2", "name": "GPT Image 2", "aliases": [], "description": "The current image model, and the migration target for every image id OpenAI has deprecated — DALL·E 2 and 3, gpt-image-1, its mini and 1.5 variants, and the chatgpt-image-latest pointer all converge here. The whole image line turned over inside eighteen months.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-oss-120b", "name": "gpt-oss-120b", "aliases": [], "description": "OpenAI's large open-weight model, released under Apache 2.0 and runnable on hardware you own. Open weights change the deprecation question entirely: OpenAI can stop hosting it, but nobody can shut down a copy you have already downloaded.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-oss-20b", "name": "gpt-oss-20b", "aliases": [], "description": "The smaller of the two gpt-oss open-weight releases, sized to run on a single consumer GPU. Like its larger sibling it carries no shutdown risk in the usual sense — the weights are Apache 2.0 licensed and already distributed.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-realtime", "name": "GPT Realtime", "aliases": [], "description": "The first realtime model to leave preview, in late 2025 — and deprecated less than a year later. Voice models have the shortest lifecycles on the platform, which matters more than usual because a realtime integration is a protocol, not just an id.", "deprecated_on": "2026-07-20", "shutdown_on": "2027-01-20", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-realtime-2.1", "recommended": true, "note": "Six months of notice, which is the most generous window in the current OpenAI deprecation wave. Use it — realtime migrations touch session handling, not just a model string.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-realtime-1.5", "name": "GPT Realtime 1.5", "aliases": [], "description": "The speech-to-speech model that replaced the gpt-4o realtime previews when those were shut down in May 2026. Realtime models use a WebSocket session rather than a request/response call, so migrating between them is a protocol change as much as an id change.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-realtime-2.1", "name": "GPT Realtime 2.1", "aliases": [], "description": "The current realtime model, named as the successor when OpenAI deprecated the entire gpt-realtime and gpt-audio generation in July 2026. It is the third realtime id in under two years.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-realtime-2.1-mini", "name": "GPT Realtime 2.1 mini", "aliases": [], "description": "The cost tier of the current realtime generation, for voice applications where per-minute audio pricing dominates the bill. It replaces gpt-realtime-mini, which is deprecated with a January 2027 shutdown.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "gpt-realtime-mini", "name": "GPT Realtime mini", "aliases": [], "description": "The cost tier of the first GA realtime generation, and the successor to gpt-4o-mini-realtime-preview. It was deprecated in July 2026, which puts anyone who migrated off the preview on their second voice migration in a year.", "deprecated_on": "2026-07-20", "shutdown_on": "2027-01-20", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-realtime-2.1-mini", "recommended": true, "note": "A same-family upgrade, so session events should carry over more cleanly than the preview-to-GA move did.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "gpt-realtime-mini-2025-10-06", "name": "GPT Realtime mini (2025-10-06)", "aliases": [], "description": "The pinned October 2025 snapshot of the realtime mini model. It was retired in July 2026, six months ahead of the floating gpt-realtime-mini id it sat behind — pinning a snapshot bought less time here, not more.", "released_on": "2025-10-06", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-realtime-2.1-mini", "recommended": true, "note": "Skip the floating gpt-realtime-mini id, which is itself deprecated with a January 2027 shutdown.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "o1", "name": "o1", "aliases": [ "o1-2024-12-17" ], "description": "The first generally available reasoning model, and the release that proved spending inference compute could substitute for a larger model. It introduced the constraints the whole o-series inherited: no system messages at launch, no streaming, and max_completion_tokens in place of max_tokens.", "released_on": "2024-12-17", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Sol lifts most of o1's restrictions — system messages, streaming and tool calls all work — so code written defensively around o1 can usually be simplified rather than ported.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "o1-mini", "name": "o1-mini", "aliases": [], "description": "The cheap reasoning model of the first o-series generation, strong on maths and code for its size but noticeably weaker on general knowledge. Retired October 2025.", "released_on": "2024-09-12", "deprecated_on": "2025-04-28", "shutdown_on": "2025-10-27", "status": "retired", "replacements": [ { "provider": "openai", "model": "o4-mini", "recommended": true, "note": "o4-mini was the named successor and is now itself deprecated, shutting down October 2026. Two migrations in eighteen months for anyone who followed the advice each time.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "o1-preview", "name": "o1-preview", "aliases": [], "description": "The first public reasoning model, released as a preview in September 2024 and retired ten months later — the shortest life of any widely adopted OpenAI model. It hid its reasoning tokens but billed for them, which is where the industry-wide argument about reasoning-token pricing started.", "released_on": "2024-09-12", "deprecated_on": "2025-04-28", "shutdown_on": "2025-07-28", "status": "retired", "replacements": [ { "provider": "openai", "model": "o3", "recommended": true, "note": "OpenAI named o3 as the successor; o3 is itself now deprecated with a December 2026 shutdown, so the durable target is gpt-5.6-sol.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "o1-pro", "name": "o1-pro", "aliases": [ "o1-pro-2025-03-19" ], "description": "The long-compute variant of o1, available only through the Responses API and priced for problems where minutes of thinking are worth it. Deprecated in April 2026 with an October shutdown.", "released_on": "2025-03-19", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "OpenAI directs o1-pro traffic to Sol with reasoning set to its pro mode rather than to a separate model id. The dial moved from the id into the request body.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "o3", "name": "o3", "aliases": [ "o3-2025-04-16" ], "description": "The reasoning model that made chain-of-thought a product feature rather than a prompting trick, and the last o-series flagship before the line was folded into GPT-5. It outlived o1 by more than a year and is still widely referenced in evaluation code and papers.", "released_on": "2025-04-16", "deprecated_on": "2026-06-11", "shutdown_on": "2026-12-11", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Both are reasoning models, so max_completion_tokens and reasoning_effort carry over. Reasoning tokens are billed as output on both, which is the line item to compare.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "o3-deep-research", "name": "o3-deep-research", "aliases": [ "o3-deep-research-2025-06-26" ], "description": "A research-agent model that ran multi-step web investigations autonomously and returned a cited report. It was only ever available on the Responses API with the web-search tool enabled, and was retired in July 2026 with its o4-mini sibling.", "released_on": "2025-06-26", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Sol plus the web-search tool covers the same ground, but you now orchestrate the research loop yourself rather than getting it built into the model.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "o3-mini", "name": "o3-mini", "aliases": [ "o3-mini-2025-01-31" ], "description": "The reasoning model that was cheap enough to use by default, and for much of 2025 the best value on the platform for anything involving maths or code. It is deprecated with an October 2026 shutdown.", "released_on": "2025-01-31", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "reasoning_effort carries over directly. If o3-mini was chosen for price rather than capability, compare Terra as well before defaulting to Sol.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "o3-pro", "name": "o3-pro", "aliases": [ "o3-pro-2025-06-10" ], "description": "The long-compute variant of o3, aimed at problems where an answer is worth minutes of thinking. It only ever ran on the Responses API, so migrations off it are usually simpler than they look — the calling code is already shaped for the successor.", "released_on": "2025-06-10", "deprecated_on": "2026-06-11", "shutdown_on": "2026-12-11", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "Raise reasoning_effort on Sol to recover o3-pro-like depth. Expect different latency characteristics: Sol returns sooner at equivalent effort than o3-pro did.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "o4-mini", "name": "o4-mini", "aliases": [ "o4-mini-2025-04-16" ], "description": "The last of the mini reasoning models, and the named successor when o1-mini was retired in 2025 — which makes it the second stop on a migration path that now needs a third. Deprecated April 2026, shutting down October 2026.", "released_on": "2025-04-16", "deprecated_on": "2026-04-22", "shutdown_on": "2026-10-23", "status": "deprecated", "replacements": [ { "provider": "openai", "model": "gpt-5.6-terra", "recommended": true, "note": "Note that OpenAI points o4-mini at Terra, not at Sol: the replacement for a small reasoning model is now a general model, not a smaller reasoner.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "deprecated" }, { "provider": "openai", "model": "o4-mini-deep-research", "name": "o4-mini-deep-research", "aliases": [ "o4-mini-deep-research-2025-06-26" ], "description": "The cheaper deep-research model, sized for research loops run at volume rather than one report at a time. Retired July 23, 2026 alongside o3-deep-research.", "released_on": "2025-06-26", "deprecated_on": "2026-04-22", "shutdown_on": "2026-07-23", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-5.6-sol", "recommended": true, "note": "No deep-research id exists in the current generation. Expect to rebuild the search-and-summarise loop in application code.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "omni-moderation", "name": "omni-moderation", "aliases": [], "description": "The multimodal moderation endpoint that replaced the text-moderation family in October 2025. It scores images as well as text, and it is free — which is why the older text-only ids had no reason to survive.", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "text-ada-001", "name": "text-ada-001", "aliases": [], "description": "The smallest InstructGPT model, fast and nearly free, used for classification and simple extraction where a wrong answer was cheap. It was the bottom rung of the ladder that ran ada, babbage, curie, davinci — a naming scheme retired along with the models themselves.", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-3.5-turbo-instruct", "recommended": true, "note": "The named successor kept the completions endpoint alive, but it is deprecated too, shutting down September 28, 2026. After that no completions-style model remains on the platform.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "text-babbage-001", "name": "text-babbage-001", "aliases": [], "description": "The second rung of the InstructGPT ladder, a step up from ada for classification work that needed slightly more nuance. Retired in January 2024 when OpenAI collapsed four base sizes into two.", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-3.5-turbo-instruct", "recommended": true, "note": "The named successor kept the completions endpoint alive, but it is deprecated too, shutting down September 28, 2026. After that no completions-style model remains on the platform.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "text-curie-001", "name": "text-curie-001", "aliases": [], "description": "The mid-tier InstructGPT model, and for a while the sweet spot on price and quality for summarisation before GPT-3.5 Turbo undercut the whole ladder at once. Retired January 2024.", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-3.5-turbo-instruct", "recommended": true, "note": "The named successor kept the completions endpoint alive, but it is deprecated too, shutting down September 28, 2026. After that no completions-style model remains on the platform.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "text-davinci-001", "name": "text-davinci-001", "aliases": [], "description": "The first instruction-tuned davinci, and the model that demonstrated that following an instruction beat completing a pattern. It was quickly overtaken by text-davinci-002 and -003, but the id survived in tutorials for years after it stopped being the right choice.", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-3.5-turbo-instruct", "recommended": true, "note": "The named successor kept the completions endpoint alive, but it is deprecated too, shutting down September 28, 2026. After that no completions-style model remains on the platform.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "text-davinci-002", "name": "text-davinci-002", "aliases": [], "description": "The first InstructGPT model trained with supervised fine-tuning on human demonstrations, and the immediate predecessor to text-davinci-003. It is the model most 2022-era research papers benchmarked against, which is why the id still turns up in code long after its January 2024 retirement.", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-3.5-turbo-instruct", "recommended": true, "note": "No current model reproduces this one's behaviour. If you are reproducing a paper rather than shipping a product, an open-weight base model is a closer comparison.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "text-davinci-003", "name": "text-davinci-003", "aliases": [], "description": "The InstructGPT model that powered the first ChatGPT preview and most of the original prompt-engineering literature. It was a completions model — you sent a prompt string and got a continuation — and its retirement in January 2024 is what forced a generation of code onto the chat API.", "released_on": "2022-11-28", "deprecated_on": "2023-07-06", "shutdown_on": "2024-01-04", "status": "retired", "replacements": [ { "provider": "openai", "model": "gpt-3.5-turbo-instruct", "recommended": true, "note": "The named successor kept the completions endpoint alive; it is now deprecated too, shutting down September 2026. After that, no completions-style model remains.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "openai", "model": "text-embedding-3-large", "name": "text-embedding-3-large", "aliases": [], "description": "The large embedding model, and the strongest retrieval quality OpenAI offers. Like its smaller sibling it is absent from the deprecations page: an embedding model cannot be retired on a chat-model schedule, because doing so would invalidate every vector index built against it.", "released_on": "2024-01-25", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "text-embedding-3-small", "name": "text-embedding-3-small", "aliases": [], "description": "The small embedding model, cheaper and stronger than the ada-002 it succeeded, with configurable output dimensions. Embedding models age differently from chat models: changing one invalidates every vector already in your index, so OpenAI has kept this line stable far longer than the chat line.", "released_on": "2024-01-25", "status": "active", "replacements": [], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "text-embedding-ada-002", "name": "text-embedding-ada-002", "aliases": [], "description": "The embedding model behind most retrieval systems built between 2022 and 2024, and one of the most commonly assumed-dead ids on the platform. It is not on OpenAI's deprecations page: re-embedding a corpus is expensive enough that the old vectors have to keep working, so this one has outlived every chat model of its era.", "released_on": "2022-12-15", "status": "active", "replacements": [ { "provider": "openai", "model": "text-embedding-3-small", "recommended": false, "note": "Not a deprecation — but 3-small is cheaper and scores better. Only worth it if you can afford to re-embed everything, since vectors from the two models are not comparable.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "active" }, { "provider": "openai", "model": "text-moderation-007", "name": "text-moderation-007", "aliases": [ "text-moderation-stable", "text-moderation-latest" ], "description": "The last of the text-only moderation classifiers, and the endpoint most trust-and-safety pipelines were built against before images became table stakes. Both floating pointers, -stable and -latest, resolved to it and went down with it in October 2025.", "deprecated_on": "2025-04-28", "shutdown_on": "2025-10-27", "status": "retired", "replacements": [ { "provider": "openai", "model": "omni-moderation", "recommended": true, "note": "The category set is wider and the response includes per-category scores for images as well as text, so any thresholds tuned against the old categories need re-checking rather than copying.", "external": false } ], "sources": [ { "url": "https://developers.openai.com/api/docs/deprecations", "title": "OpenAI API deprecations", "accessed": "2026-07-30" } ], "last_verified": "2026-07-30", "computed_status": "retired" }, { "provider": "qwen", "model": "gte-rerank", "name": "GTE Rerank", "aliases": [], "description": "The reranker Alibaba inherited from its GTE embedding research line, and the last Model Studio id that was not Qwen-branded. It was retired on 30 May 2026, which completed the rebranding of every retrieval model on the platform.", "shutdown_on": "2026-05-30", "status": "retired", "replacements": [ { "provider": "qwen", "model": "qwen3-rerank", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "qwen", "model": "qwen-flash-2025-07-28", "name": "Qwen Flash (2025-07-28)", "aliases": [], "description": "The replacement Alibaba named for the retired qwen-turbo line, and the point at which turbo stopped being a tier name in Model Studio. Anything still calling qwen-turbo with a 2024 date stamp was pointed here.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "qwen", "model": "qwen-plus-2024-11-27", "name": "Qwen Plus (2024-11-27)", "aliases": [], "description": "The November 2024 Plus snapshot, decommissioned on 30 January 2026 after fourteen months. The qwen-plus family name has outlived four model generations by rotating dated snapshots underneath it, which is why the successor is another qwen-plus rather than a qwen3 id.", "shutdown_on": "2026-01-30", "status": "retired", "replacements": [ { "provider": "qwen", "model": "qwen-plus-2025-12-01", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "qwen", "model": "qwen-plus-2025-12-01", "name": "Qwen Plus (2025-12-01)", "aliases": [], "description": "The December 2025 snapshot of the long-running qwen-plus line, and the named replacement for the 2024-11-27 snapshot that was deprecated in January 2026. Alibaba keeps qwen-plus alive across generations as a dated series rather than retiring the family name, which is why this id looks nothing like qwen3.7-plus.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "qwen", "model": "qwen-turbo-2024-09-19", "name": "Qwen Turbo (2024-09-19)", "aliases": [], "description": "The last qwen-turbo snapshot. Its decommissioning is where the turbo tier name disappeared from Model Studio: Alibaba pointed it at qwen-flash rather than at another turbo, and no turbo id has shipped since.", "shutdown_on": "2026-01-30", "status": "retired", "replacements": [ { "provider": "qwen", "model": "qwen-flash-2025-07-28", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "qwen", "model": "qwen3-14b", "name": "Qwen3 14B", "aliases": [], "description": "The 14B dense Qwen3. Alibaba points every retired dense model at qwen3.6-flash regardless of the original size, which for this one is a move from a mid-size dense model to a much larger sparse one - cheaper per token, but not the same thing at all.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.6-flash", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-30b-a3b", "name": "Qwen3 30B A3B", "aliases": [], "description": "The 30B mixture of experts with 3B active, the most-downloaded model of the Qwen3 generation because it ran fast on consumer hardware. It is the only dense-line retirement Alibaba routes to Plus rather than Flash.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.7-plus", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-32b", "name": "Qwen3 32B", "aliases": [], "description": "The largest hosted dense Qwen3, and the last of the four to be scheduled. Its retirement completes Alibaba's move away from serving open dense checkpoints as paid API ids - the weights stay on Hugging Face, the endpoints do not.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.6-flash", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-8b", "name": "Qwen3 8B", "aliases": [], "description": "One of the dense Qwen3 models Alibaba hosted alongside the open-weights release, for people who wanted the exact 8B model without running it themselves. All four dense sizes on Model Studio go on the same day, ending the hosted dense line entirely.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.6-flash", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-coder-plus", "name": "Qwen3 Coder Plus", "aliases": [], "description": "Alibaba's dedicated coding model, popular well outside China because the open Qwen3-Coder weights made it easy to try before paying for the hosted id. Its retirement folds coding back into the general Plus tier, the same consolidation Mistral and Moonshot made in the same period.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.7-plus", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-max", "name": "Qwen3 Max", "aliases": [], "description": "The stable Max id of the third generation, still served while both of its dated snapshots are scheduled to go on 10 October 2026. It is the named successor to qwen3-max-preview, which is the ordinary shape of an Alibaba retirement: the preview dies, the stable id it previewed lives on.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "qwen", "model": "qwen3-max-2025-09-23", "name": "Qwen3 Max (2025-09-23)", "aliases": [], "description": "The launch snapshot of Qwen3 Max from September 2025. It survives its own successor snapshot's announcement by exactly nothing - Alibaba set both to go on 10 October 2026, so pinning the older one bought no extra time.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.7-max", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-max-2026-01-23", "name": "Qwen3 Max (2026-01-23)", "aliases": [], "description": "The January 2026 Max snapshot, five months old at the point Alibaba scheduled it for decommissioning. Dated Qwen snapshots get thirty days' notice against three months for stable ids, which is the argument for never pinning one unless you are reproducing an evaluation.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.7-max", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-max-preview", "name": "Qwen3 Max Preview", "aliases": [], "description": "The preview that introduced the Qwen3 Max tier in 2025, kept alive long after the stable qwen3-max shipped because so much early integration code pinned it. It is the textbook Alibaba retirement: the preview goes, the stable id it previewed stays.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3-max", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-rerank", "name": "Qwen3 Rerank", "aliases": [], "description": "The reranker that replaced gte-rerank when Alibaba retired the GTE line in May 2026. It is the only survivor of Model Studio's pre-Qwen-branded retrieval models, all of which have now been folded into the Qwen naming.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "qwen", "model": "qwen3-vl-8b-instruct", "name": "Qwen3 VL 8B Instruct", "aliases": [], "description": "The small instruction-tuned vision model, 8B and open-weighted, which made it the usual choice for self-hosted image understanding in the Qwen ecosystem. The hosted id goes on 10 October 2026; the weights do not.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.6-flash", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-vl-8b-thinking", "name": "Qwen3 VL 8B Thinking", "aliases": [], "description": "The reasoning variant of the 8B vision model, shipped as a separate id rather than a parameter. Splitting thinking and non-thinking into two model strings is the convention every major lab abandoned during 2026, and Alibaba retiring both halves on one day is the end of it here.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.6-flash", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-vl-flash", "name": "Qwen3 VL Flash", "aliases": [], "description": "The stable vision-language Flash id, and the head of a family of four VL models all retiring on the same day. Alibaba is not replacing the VL line with a newer VL model: vision moved into the general Plus and Flash models, so the whole qwen3-vl namespace disappears.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.6-flash", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-vl-flash-2025-10-15", "name": "Qwen3 VL Flash (2025-10-15)", "aliases": [], "description": "The original VL Flash snapshot from October 2025. Both dated VL Flash snapshots and the undated stable id retire together, which means there is no version of this model left to fall back to.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.6-flash", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3-vl-flash-2026-01-22", "name": "Qwen3 VL Flash (2026-01-22)", "aliases": [], "description": "The January 2026 snapshot of VL Flash, and the newest model in Alibaba's entire decommissioning table - under nine months from release to scheduled shutdown.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.6-flash", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3.5-omni-plus", "name": "Qwen3.5 Omni Plus", "aliases": [], "description": "The omni-modal model, handling text, audio, image and video in and speech out. Alibaba versions Omni on its own track - it is still at 3.5 while the text line has moved to 3.7 - so the two families' lifecycles are unrelated.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "qwen", "model": "qwen3.5-omni-plus-realtime", "name": "Qwen3.5 Omni Plus Realtime", "aliases": [], "description": "The streaming variant of Omni Plus, for live voice conversation rather than batched audio. It is a separate id from the non-realtime model, which means a voice integration carries its own deprecation risk independent of the batch one.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "qwen", "model": "qwen3.6-flash", "name": "Qwen3.6 Flash", "aliases": [], "description": "The cheap fast model of the current range, and the single largest destination in Alibaba's deprecation table: eight retiring ids point here, including the entire qwen3 dense series and most of the VL line. Note that Model Studio's own decommissioning page also lists qwen3.6-flash as retiring with itself named as the replacement, which is a documentation error rather than a real retirement.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "qwen", "model": "qwen3.6-max-preview", "name": "Qwen3.6 Max Preview", "aliases": [], "description": "A preview of the 3.6 Max that Alibaba superseded before it ever went stable - there is no qwen3.6-max, only this preview and then 3.7. Preview ids in Model Studio get the same thirty-day notice as any other snapshot, which is not much for something people treated as a flagship.", "shutdown_on": "2026-10-10", "status": "deprecated", "replacements": [ { "provider": "qwen", "model": "qwen3.7-max", "recommended": true, "external": false } ], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/model-depreciation", "title": "Alibaba Cloud Model Studio model decommissioning", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "deprecated" }, { "provider": "qwen", "model": "qwen3.7-max", "name": "Qwen3.7 Max", "aliases": [], "description": "Alibaba's current flagship and the migration target for every retired Max snapshot as well as the qwen3-coder-plus line. Model Studio gives stable ids three months' notice and dated snapshots thirty days, so an id like this one without a date stamp is the safer thing to pin.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "qwen", "model": "qwen3.7-plus", "name": "Qwen3.7 Plus", "aliases": [], "description": "The mid tier of the 3.7 generation, and now also the vision model - Alibaba folded image and video understanding into Plus rather than keeping a separate qwen3-vl line, which is why the whole VL family is scheduled for retirement.", "status": "active", "replacements": [], "sources": [ { "url": "https://www.alibabacloud.com/help/en/model-studio/models", "title": "Alibaba Cloud Model Studio models", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-2-1212", "name": "Grok 2", "aliases": [ "grok-2", "grok-2-latest" ], "description": "The model that took xAI's API out of beta in December 2024, replacing grok-beta with better instruction-following and multilingual coverage at $2 per million input tokens. It is gone: the id is absent from xAI's model list and its documentation page returns a 404. xAI never published a retirement date for it, and this catalogue will not invent one.", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": false, "note": "Our suggestion, not xAI's - xAI removed the model's page rather than publishing a migration target for it.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing (no longer lists this model)", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-2-image-1212", "name": "Grok 2 Image", "aliases": [], "description": "The Aurora-based image generator that xAI ran before the Imagine line, priced per image rather than per token and capped at four images a request. It is no longer served and its docs page is gone. xAI announced the shutdown by email and in-product rather than in the documentation, so there is no citable first-party date left.", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-imagine-image", "recommended": false, "note": "The Imagine line is the direct successor, but xAI's public docs no longer carry the notice that said so.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing (no longer lists this model)", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-2-vision-1212", "name": "Grok 2 Vision", "aliases": [], "description": "xAI's first image-understanding model on the API, shipped alongside grok-2-1212 and limited to 8k of context, which made it awkward for anything beyond a single image and a short question. Its documentation page now 404s and the id is off the model list; no retirement date was ever published in xAI's docs.", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": false, "note": "grok-4.3 accepts images natively with a 1M window. This is our reading of the path, not a migration xAI documented.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing (no longer lists this model)", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-3", "name": "Grok 3", "aliases": [], "description": "The model xAI trained on the Colossus cluster and launched in February 2025 with a Super Bowl-scale marketing push, reaching the API that April. It is the oldest id in the May 2026 retirement wave and the only one whose replacement route runs through a non-reasoning setting, since Grok 3 never thought by default.", "shutdown_on": "2026-05-15", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": true, "note": "Maps to grok-4.3 with reasoning effort \"none\".", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-3-mini", "name": "Grok 3 Mini", "aliases": [], "description": "The small reasoning model of the Grok 3 generation, and the first xAI id that exposed its thinking trace, which made it briefly popular for debugging agent behaviour. It went before grok-3 itself did. The id is off the model list and its docs page 404s, with no retirement date left in xAI's documentation.", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": false, "note": "grok-4.3 with a low reasoning effort is the closest equivalent, though the price floor is considerably higher than Grok 3 Mini's was.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing (no longer lists this model)", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-4-0709", "name": "Grok 4", "aliases": [ "grok-4", "grok-4-latest" ], "description": "The July 2025 flagship that xAI launched with claims of PhD-level performance across every subject, and the first Grok that could not be asked to skip thinking. Because the undated grok-4 alias pointed here, plenty of integrations were running it without ever naming the snapshot.", "shutdown_on": "2026-05-15", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": true, "note": "Maps to grok-4.3 with reasoning effort \"low\". Grok 4 had no non-reasoning mode at all, so this is the first time the workload can be run without thinking tokens if you want that.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-4-1-fast-non-reasoning", "name": "Grok 4.1 Fast Non-Reasoning", "aliases": [], "description": "The straight-through twin of Grok 4.1 Fast Reasoning, same weights and same 2M window with thinking turned off. Splitting one model into two ids by reasoning mode is exactly what grok-4.3's reasoning_effort parameter made unnecessary, and both halves went at once.", "shutdown_on": "2026-05-15", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": true, "note": "Maps to grok-4.3 with reasoning effort \"none\". Set that explicitly rather than relying on the redirect, or a later change to the default leaves you paying for thinking tokens you never asked for.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-4-1-fast-reasoning", "name": "Grok 4.1 Fast Reasoning", "aliases": [], "description": "A 2M-context reasoning model that was, for a few months, the cheapest way to put a long document in front of a frontier model. It retired on 15 May 2026, and the slug still resolves - requests are silently served by grok-4.3 at grok-4.3 prices, which is the trap in this whole retirement wave.", "shutdown_on": "2026-05-15", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": true, "note": "xAI maps this id to grok-4.3 with reasoning effort \"low\". The slug keeps working, so nothing breaks - but you are billed $1.25/$2.50 per million instead of the old rate, and the context window drops from 2M to 1M.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-4-fast-non-reasoning", "name": "Grok 4 Fast Non-Reasoning", "aliases": [], "description": "The non-thinking half of Grok 4 Fast, and the cheapest id xAI has ever served. It was the natural home for classification and extraction jobs at volume, which makes it the one in this wave most likely to be sitting in a batch pipeline nobody has looked at since it was written.", "shutdown_on": "2026-05-15", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": true, "note": "Maps to grok-4.3 with reasoning effort \"none\". This is the largest price jump in the wave: a high-volume extraction job that cost pennies on this id is now billed at flagship rates.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-4-fast-reasoning", "name": "Grok 4 Fast Reasoning", "aliases": [], "description": "The September 2025 model that introduced the 2M context window to the xAI API and undercut every comparable frontier model on price, which is why it ended up in so many agent frameworks as the default. Retired 15 May 2026, one generation behind 4.1 Fast and shut down on the same day as it.", "shutdown_on": "2026-05-15", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": true, "note": "Maps to grok-4.3 with reasoning effort \"low\".", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-4.20-0309-non-reasoning", "name": "Grok 4.20 Non-Reasoning", "aliases": [], "description": "The straight-through half of the 4.20 snapshot pair, for workloads that want a 1M-token window without paying for thinking tokens. Every previous xAI id with non-reasoning in the name was retired on 15 May 2026 in favour of a reasoning_effort parameter, so the naming here is a step back to an older convention.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-4.20-0309-reasoning", "name": "Grok 4.20 Reasoning", "aliases": [], "description": "A dated 4.20 snapshot that keeps reasoning and non-reasoning as separate ids, the convention xAI had just moved away from with 4.3. Pinned snapshots like this one are the safe choice for evaluation harnesses, at the cost of being first in line whenever xAI runs its next retirement wave.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-4.20-multi-agent-0309", "name": "Grok 4.20 Multi-Agent", "aliases": [], "description": "xAI's heavy tier, running several agents over the same problem and reconciling their answers inside one API call. It is the descendant of Grok 4 Heavy, which never got an API id at all - this is the first time that mode has been callable rather than being a consumer subscription feature.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-4.3", "name": "Grok 4.3", "aliases": [], "description": "The workhorse of the xAI API and the destination for seven of the eight ids retired on 15 May 2026. It takes a reasoning_effort setting rather than splitting reasoning and non-reasoning into separate model names, which is the change that made half of xAI's catalogue redundant in one release. A million tokens of context at $1.25 per million in.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" }, { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-4.5", "name": "Grok 4.5", "aliases": [ "grok-4.5-latest" ], "description": "xAI's current flagship, with a 500k context window and a February 2026 knowledge cutoff. It is worth noting that it is not the model xAI routes retired ids to - that is grok-4.3, which has double the context and costs less. 4.5 is a deliberate choice rather than a default you fall into.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-beta", "name": "Grok Beta", "aliases": [], "description": "The id xAI's public API launched with in November 2024, during the period when the whole platform was free credit and a waitlist. It lasted about a month before grok-2-1212 superseded it. Nothing in xAI's current documentation records when it stopped answering.", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": false, "note": "Two full generations later. Treat this as a rewrite rather than a model swap.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing (no longer lists this model)", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-build-0.1", "name": "Grok Build 0.1", "aliases": [], "description": "The coding model that replaced grok-code-fast-1 when it retired in May 2026, and the only id in xAI's range with a 0.x version number. Its 256k context is the smallest xAI serves, which is a deliberate trade for the throughput an editor integration needs.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-code-fast-1", "name": "Grok Code Fast 1", "aliases": [], "description": "xAI's first dedicated coding model, given away free through GitHub Copilot and Cursor for long enough to build real habits around it. It is the only id in the May 2026 wave that does not route to grok-4.3 - coding traffic goes to grok-build-0.1 instead.", "shutdown_on": "2026-05-15", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-build-0.1", "recommended": true, "note": "The context window shrinks from 256k on both sides, but grok-build-0.1 is a different model rather than a tuned reasoning effort, so agent prompts are worth re-testing rather than lifting straight across.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-imagine-image", "name": "Grok Imagine Image", "aliases": [], "description": "xAI's image generation endpoint and the successor to the grok-2-image line. The Imagine family renamed everything: nothing in xAI's current image or video range carries a Grok version number any more, which makes the lineage from the 1212-era models harder to trace than it should be.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-imagine-image-pro", "name": "Grok Imagine Image Pro", "aliases": [], "description": "The premium tier of xAI's image generator, retired in the same wave as seven text models and renamed rather than replaced. Nothing about the model changed on 15 May 2026; the word \"pro\" was simply swapped for \"quality\" in the id.", "shutdown_on": "2026-05-15", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-imagine-image-quality", "recommended": true, "note": "A rename. Unlike the text models in this wave there is no price or context change to plan around - just the new string.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-imagine-image-quality", "name": "Grok Imagine Image Quality", "aliases": [], "description": "The slower, higher-fidelity image model, and the id that grok-imagine-image-pro was folded into on 15 May 2026. It is the only non-text model in that retirement wave, and the only one that did not end up pointing at grok-4.3.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" }, { "url": "https://docs.x.ai/developers/migration/may-15-retirement", "title": "Grok Model Retirement on May 15, 2026", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-imagine-video", "name": "Grok Imagine Video", "aliases": [], "description": "Text and image to video generation over the same API surface as the rest of Grok, which is unusual - most providers keep video behind a separate async jobs endpoint. It is the base tier of a two-model video line that xAI has already versioned once.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-imagine-video-1.5", "name": "Grok Imagine Video 1.5", "aliases": [], "description": "The newer video model, shipped as a separate id rather than as an upgrade to grok-imagine-video. Both are still listed, which means xAI is running the two in parallel and has not yet said what happens to the unversioned one - the same pattern that preceded the May 2026 image retirement.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-vision-beta", "name": "Grok Vision Beta", "aliases": [], "description": "The image-input companion to grok-beta from the November 2024 API launch, and the shortest-lived id xAI has shipped - grok-2-vision-1212 replaced it within weeks. Like the rest of the beta-era models it left no dated notice behind in the docs.", "status": "retired", "replacements": [ { "provider": "xai", "model": "grok-4.3", "recommended": false, "note": "Our suggestion. xAI published no migration path for the beta ids.", "external": false } ], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing (no longer lists this model)", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "retired" }, { "provider": "xai", "model": "grok-voice-think-fast-1.0", "name": "Grok Voice Think Fast 1.0", "aliases": [], "description": "A speech-to-speech model tuned for conversational latency, the audio half of xAI's range. The \"think fast\" naming is doing real work: it is a reasoning model with the budget capped low enough that a human on a phone call does not notice the pause.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xai", "model": "grok-voice-think-fast-2.0", "name": "Grok Voice Think Fast 2.0", "aliases": [], "description": "The second generation of xAI's voice model, listed alongside 1.0 rather than replacing it. xAI has not published a deprecation notice for the 1.0 id, so both remain callable - but on this provider's record, the older of two co-listed ids is the one that disappears next.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.x.ai/developers/models", "title": "xAI models and pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xiaomi", "model": "mimo-v2.5", "name": "MiMo V2.5", "aliases": [], "description": "The base omni-modal model, taking text, image, video and audio with a million tokens of context. It replaced the MiMo V2 series, which Xiaomi deprecated on 30 June without publishing individual model ids for the models going away - so this catalogue records the successor and not the predecessors.", "status": "active", "replacements": [], "sources": [ { "url": "https://mimo.mi.com/docs/en-US/quick-start/model-hyperparameters", "title": "Xiaomi MiMo model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xiaomi", "model": "mimo-v2.5-asr", "name": "MiMo V2.5 ASR", "aliases": [], "description": "The speech recognition model, bilingual across Chinese and English including dialects, and tuned for noisy input. Xiaomi splitting ASR out from the omni-modal model is a hint about where these run: on devices, where a dedicated small model beats a general one.", "status": "active", "replacements": [], "sources": [ { "url": "https://mimo.mi.com/docs/en-US/quick-start/model-hyperparameters", "title": "Xiaomi MiMo model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xiaomi", "model": "mimo-v2.5-pro", "name": "MiMo V2.5 Pro", "aliases": [], "description": "Xiaomi's flagship, a trillion-parameter model with 42B active and a million-token window, released open-weights under MIT. A phone manufacturer shipping a frontier model under a permissive licence is the unusual part: the hosted API is a convenience, not the only way to run it.", "status": "active", "replacements": [], "sources": [ { "url": "https://mimo.mi.com/docs/en-US/quick-start/model-hyperparameters", "title": "Xiaomi MiMo model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "xiaomi", "model": "mimo-v2.5-pro-ultraspeed", "name": "MiMo V2.5 Pro UltraSpeed", "aliases": [], "description": "The same Pro weights served with FP4 quantisation and parallel decoding for peaks around a thousand tokens a second. It is a separate model id rather than a serving flag, so a latency choice is baked into the string your code sends and carries its own lifecycle.", "status": "active", "replacements": [], "sources": [ { "url": "https://mimo.mi.com/docs/en-US/quick-start/model-hyperparameters", "title": "Xiaomi MiMo model list", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4-32b-0414-128k", "name": "GLM-4-32B-0414-128K", "aliases": [], "description": "The last survivor of the GLM-4 generation, a 32B model from April 2024 that predates the 4.5 rewrite and everything after it. Its date-stamped id belongs to an older naming scheme Z.ai abandoned; it is the only model in the current pricing table that still uses one.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.5", "name": "GLM-4.5", "aliases": [], "description": "355B total parameters with 32B active, released July 2025 under an open licence, and the model that established Z.ai as a serious lab outside China. It is the oldest GLM still listed for sale, a year and five generations later.", "released_on": "2025-07-28", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.5-air", "name": "GLM-4.5-Air", "aliases": [], "description": "106B total parameters with 12B active, the lightweight GLM-4.5 and by some distance the most-downloaded model Z.ai has released. It is the id most likely to be embedded in someone else's product, which makes its complete absence from Z.ai's deprecation communications worth noting.", "released_on": "2025-07-28", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.5-airx", "name": "GLM-4.5-AirX", "aliases": [], "description": "The fast-serving Air, combining both of the suffixes Z.ai used in the 4.5 generation. Four ids for one model family - base, X, Air and AirX - is the high water mark of Z.ai's naming, and the 4.6 line dropped straight back to one.", "released_on": "2025-07-28", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.5-flash", "name": "GLM-4.5-Flash", "aliases": [], "description": "The free member of the GLM-4.5 family, and the entry point most people used to try Z.ai at all. A year on it is still listed at the same price, which is to say none, alongside the newer GLM-4.7-Flash that was meant to succeed it.", "released_on": "2025-07-28", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.5-x", "name": "GLM-4.5-X", "aliases": [], "description": "The high-performance tier of the GLM-4.5 launch, tuned for throughput at the top of the range. The X suffix was retired as a naming convention after 4.5 - nothing in the 4.6, 4.7 or 5.x lines uses it - but the ids themselves are still served.", "released_on": "2025-07-28", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.5v", "name": "GLM-4.5V", "aliases": [], "description": "Z.ai's first widely-used vision model, released August 2025 on the GLM-4.5-Air base and notable for reading UI screenshots well enough to drive them. It is the oldest vision id still listed and the direct ancestor of everything in the V-series since.", "released_on": "2025-08-11", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.6", "name": "GLM-4.6", "aliases": [], "description": "The model that took Z.ai's context window from 128k to 200k and became the default open-weights choice for coding agents through late 2025. It is named in the GLM-5.2 migration guide as a model to move off, but Z.ai has published no date for it and it is still on the pricing page.", "released_on": "2025-09-30", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.6v", "name": "GLM-4.6V", "aliases": [], "description": "The December 2025 vision model, released two months after the text GLM-4.6 it is built on. It anchors a three-model tier of its own - base, FlashX and Flash - the same pricing split Z.ai applies to its text models.", "released_on": "2025-12-08", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.6v-flash", "name": "GLM-4.6V-Flash", "aliases": [], "description": "The free vision tier. Between GLM-4.5-Flash, GLM-4.7-Flash and this, Z.ai runs three no-cost ids simultaneously - a strategy that gets the models adopted and also means a lot of production code is sitting on endpoints with no support commitment attached.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.6v-flashx", "name": "GLM-4.6V-FlashX", "aliases": [], "description": "The cheap fast tier of GLM-4.6V, for image workloads at volume where per-request cost matters more than the last few points of accuracy. Vision requests are token-expensive, which is why Z.ai bothers to run three price points for one vision model.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.7", "name": "GLM-4.7", "aliases": [], "description": "The December 2025 model that closed out the 4.x line, with large gains in multilingual agentic coding and terminal work over GLM-4.6. It is still one of the three models supported on every Z.ai coding plan, which makes it the oldest id the company actively promotes.", "released_on": "2025-12-22", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.7-flash", "name": "GLM-4.7 Flash", "aliases": [], "description": "The free tier of the GLM-4.7 family, released a month after the base model. Z.ai has shipped a Flash id in every generation since 4.5, and they are the ids most likely to be running in hobby projects and evaluation scripts that nobody is watching for breakage.", "released_on": "2026-01-19", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-4.7-flashx", "name": "GLM-4.7 FlashX", "aliases": [], "description": "The mid-tier speed variant of GLM-4.7. Z.ai runs a three-way split - base, Flash and FlashX - where Flash is free or near-free, FlashX is cheap and fast, and the base model is neither. The suffixes carry pricing information rather than capability information.", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-5", "name": "GLM-5", "aliases": [], "description": "The February 2026 model that started Z.ai's fifth generation, superseded twice within four months. Several downstream inference providers retired their GLM-5 endpoints in April 2026; Z.ai itself published nothing, and the model remains on its pricing page.", "released_on": "2026-02-12", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-5-turbo", "name": "GLM-5 Turbo", "aliases": [], "description": "The fast, cheap member of the GLM-5 family, and one of only three models available across every tier of Z.ai's coding plan. Turbo here means a smaller model rather than the same weights served faster, which is the opposite of what the suffix means at MiniMax or Moonshot.", "released_on": "2026-03-15", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-5.1", "name": "GLM-5.1", "aliases": [], "description": "The April 2026 refresh of GLM-5, and the model people were pushed to when third-party hosts started dropping GLM-5. On Z.ai's own platform both are still listed and priced, which is the difference worth knowing: a model can be deprecated on someone else's inference service while the source provider keeps serving it.", "released_on": "2026-04-07", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-5.2", "name": "GLM-5.2", "aliases": [], "description": "Z.ai's current flagship, released June 2026 to unify frontier reasoning, coding and agentic work in one model. It has its own migration guide covering GLM-5.1, GLM-5, GLM-4.7, GLM-4.6 and GLM-4.5 - a document that reads like a deprecation notice but names no dates, which is Z.ai's pattern throughout.", "released_on": "2026-06-16", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-5v-turbo", "name": "GLM-5V-Turbo", "aliases": [], "description": "The vision model of the GLM-5 generation, and the only one - there is no base GLM-5V, just the turbo variant. Z.ai's V-series numbering tracks the text models one release behind, so the vision line is always a generation adrift of the flagship.", "released_on": "2026-04-01", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-asr-2512", "name": "GLM-ASR-2512", "aliases": [], "description": "The speech recognition model, and the only current Z.ai id with a date stamp baked into the name. That stamp is the one piece of lifecycle signalling Z.ai gives anywhere: it implies a 2512 successor is expected, which is more than the undated ids tell you.", "released_on": "2025-12-10", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-image", "name": "GLM-Image", "aliases": [], "description": "Z.ai's image generation model, released January 2026 and unversioned since. It carries no number at all, unlike every other id in the range, which leaves no obvious way to tell whether a future release replaces it or sits beside it.", "released_on": "2026-01-14", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" }, { "provider": "zai", "model": "glm-ocr", "name": "GLM-OCR", "aliases": [], "description": "Z.ai's document-parsing model, split out from the general vision line in February 2026. Keeping OCR separate from vision is the same call Mistral made and the opposite of Google's - the argument being that structured document extraction wants different training data than general image understanding.", "released_on": "2026-02-03", "status": "active", "replacements": [], "sources": [ { "url": "https://docs.z.ai/release-notes/new-released", "title": "Z.ai new releases", "accessed": "2026-08-03" }, { "url": "https://docs.z.ai/guides/overview/pricing", "title": "Z.ai model pricing", "accessed": "2026-08-03" } ], "last_verified": "2026-08-03", "computed_status": "active" } ] }