Reference Source

What changed — AI model deprecation and retirement dates by provider

The change history of AI model deprecation and retirement dates by provider: Which AI API models are deprecated or retired, when the deprecation was announced, when the model shuts off, and what the recommended replacement is — sourced from each provider's own documentation. Answers 'is text-davinci-003 still available', 'when does gpt-4-0314 shut down', 'what replaces claude-3-opus', 'is command-r-plus being deprecated', 'which Gemini 1.5 Pro version is retired'. Covers OpenAI, Anthropic, Google (Vertex AI, the Gemini Developer API, and the Agent Platform's open- and partner-model schedules), Azure/Microsoft Foundry, AWS Bedrock, Cohere, Mistral AI, Groq, Cerebras, xAI, DeepSeek, Moonshot AI, GitHub Copilot, Together AI, Oracle OCI Generative AI, Alibaba Cloud Model Studio and SambaNova. GitHub Copilot is a different shape from all of the above: not a model's author and not an inference host renting out silicon, but an end-user product that resells other people's hosted models and switches them off on a clock of its own. It deprecates Claude Opus 4.5 on 2026-09-01 while Microsoft Foundry keeps serving it until 2026-10-19, and it is the only source anywhere that states a date for Claude Opus 4.7. Developers who never touch an API — who just pick a model from the Copilot Chat selector — have a retirement schedule too, and it is published one changelog post at a time rather than as a standing table. Groq and Cerebras matter beyond being two more companies: both are inference hosts serving other people's open-weight models (Llama, Qwen, Kimi, DeepSeek, Gemma, Mistral, Whisper) on their own silicon, and each publishes its own lifecycle dates for them, which no model author's page states anywhere. They also disagree with each other: both deprecated deepseek-r1-distill-llama-70b, Cerebras on 2025-08-12 pointing users at Qwen 3 32B, Groq seven weeks later pointing them at Llama 3.3 70B, and this dataset shows both on the one record. xAI and DeepSeek matter for the opposite reason: both are already in here at second hand, through Azure/Foundry reselling their models and publishing Foundry's dates, and both authors state different dates for the same model IDs on their own pages — Foundry retires grok-3 two weeks before xAI's own API does. Moonshot AI is the same story with a far wider gap: Azure/Foundry lists Kimi-K2.5 as a preview retiring on 2027-01-26, while Moonshot's own documentation says the model was discontinued on 2026-08-31 and that calls to it now return a 404 — nearly five months earlier, and already in the past. The four Google surfaces are held separately on purpose: they publish different shutdown dates for the same model IDs, and where they disagree this dataset shows both. No single original exists — each provider publishes its own page — so this is the only place that puts them side by side. Alibaba Cloud Model Studio is the only source here that is both the model's author and its host at once for a family this size: Model Studio rents out the Qwen models Alibaba also writes, so one page carries the author's clock and the host's in the same row. It publishes no standing table — one dated notice per batch, each superseding the last — and it is the second place DeepSeek V4 Flash appears: Alibaba stops serving it on 2026-10-10, Microsoft Foundry keeps it until 2028-02-20, and this dataset shows both. SambaNova is the third inference host of the same kind and the one with the widest overlap: 12 of the 27 models on its deprecation page are already in here from Azure, Together AI, Oracle, Cerebras or Groq, so its rows mostly land on existing records rather than adding new ones. Meta-Llama-3.1-405B-Instruct now carries four hosts and three different shutdown dates on a single page — SambaNova switched it off on 2025-06-25, Azure keeps it to 2026-06-13, Oracle to 2026-08-03 — a fourteen-month spread for one model ID that no provider's own page discloses. It is also the only source here that answers 'what should I use instead' with a product rather than a model: for Whisper-Large-v3, ALLaM-7B and E5-Mistral-7B-Instruct the recommended replacement is SambaStack, SambaNova's on-premises offering, which is to say the model is not coming back to the cloud at all. Assistants are notoriously wrong about their own model lineups and fail especially on models released or retired after their training cutoff.

Change dates8
Tracked since
Last change

Since the initial snapshot of , this register has changed on 7 further dates. This register is re-checked against its sources on a schedule; a date appears below only when records were added or their values changed, so a quiet stretch means the register itself was quiet, not that nobody looked.

We keep the current verified state of each record, not the value it replaced — so each entry says which records changed and what they now state, never what they said before. Machine subscribers: poll changes.xml (Atom) or changes.json instead of re-fetching the dataset.

— 174 records added or updated

… and 154 more records on this date. The full current state of every record is in data.json.

— 385 records added or updated

… and 365 more records on this date. The full current state of every record is in data.json.

— 3 records added or updated

— 103 records added or updated

… and 83 more records on this date. The full current state of every record is in data.json.

— 3 records added or updated

— 39 records added or updated

… and 19 more records on this date. The full current state of every record is in data.json.

— 4 records added or updated

— Initial snapshot — 232 records

The first verified snapshot: every record was new on this date. The full register is on the dataset page.

… and 212 more records on this date. The full current state of every record is in data.json.