On 30 November, Anthropic retires Claude Sonnet 4.5 and requests to it start failing. On 1 December, Google stops serving Gemini 2.5 Pro — but your requests will not fail. They will be answered by a different model, billed at that model’s rates, with no error and no interruption.
That difference is the whole point of this article. Three major platforms are retiring models over the next five months, and they have settled on two opposite ways of doing it. One breaks loudly. The other changes your output and your invoice while everything appears to be working.
Here is what is being switched off, when, and what each platform does to the requests that are still pointing at it.
The calendar
| Date | What happens | Platform |
|---|---|---|
| 30 Nov 2026 | claude-sonnet-4-5 retires — requests fail | Anthropic |
| 1 Dec 2026 | gemini-2.5-pro and gemini-2.5-flash start re-routing | |
| 3 Dec 2026 | gemini-omni-1.1-flash starts re-routing | |
| 6 Jan 2027 | tts-1, tts-1-hd and gpt-4o-mini-tts shut down | OpenAI |
| 14 Jan 2027 | GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down | OpenAI |
| 20 Jan 2027 | Legacy audio and realtime models shut down | OpenAI |
| 1 Feb 2027 | gpt-5.5-cyber-preview shuts down | OpenAI |
| 10 Feb 2027 | gpt-image-2 shuts down | OpenAI |
| 26 Feb 2027 | whisper-1 and the 4o transcription family shut down | OpenAI |
| 1 Apr 2027 | gpt-5.1, gpt-5.3-codex and gpt-5.4-nano shut down | OpenAI |
Google: nothing breaks, which is the problem
Google’s release notes are unusually direct about what happens. On Gemini 2.5 Pro: “Effective December 1, 2026, all API requests to gemini-2.5-pro will automatically route to gemini-3.8-flash and be billed at Gemini 3.8 Flash rates.”
Read that twice. A Pro-class model is being replaced, silently, by a Flash-class one. Flash models are built for speed and cost rather than depth. If your application was using 2.5 Pro because it needed careful reasoning, the quality of what comes back on 1 December will change, and nothing in your logs will say so.
The same thing happens one rung down: gemini-2.5-flash routes to gemini-3.5-flash-lite “and be billed at Gemini 3.5 Flash-Lite rates”. Two days later, on 3 December, gemini-omni-1.1-flash routes to gemini-omni-1.2-flash.
Google also lists what will actually break if you do nothing, and it is not nothing: “Breaking API changes include function calling and thought_signature validation, thinking_level replacing thinking_budget, and removed sampling parameters and prefilled model turns.”
So the honest summary of the Google deadline is: your request will be answered, your bill will change, and if you use function calling or tuned sampling parameters, parts of your integration will fail anyway — just not the part that would have told you a model had been retired.
Not every Google re-route is a downgrade. The embeddings one is explicitly benign: gemini-embedding-2-preview now routes to gemini-embedding-2, and Google notes it “uses the same API schema, pricing, and embedding space, so no other code changes or vector re-indexing are needed.” That sentence is there because, for embeddings, a silent model swap would otherwise mean your stored vectors no longer match your new ones — a far nastier failure than an error message.
OpenAI: the longest notice, and the largest cull
OpenAI is the only one of the three that publishes a formal notice policy, and it is a good one:
- Generally available models: at least 6 months.
- Specialised variants and image generation models: at least 3 months.
- Preview models: as little as 2 weeks. OpenAI’s own warning is worth repeating — “We don’t recommend using preview models for business-critical production workloads unless you can migrate on short notice.”
The headline date is 14 January 2027, when the GPT-4 series goes. GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini all shut down that day, with gpt-6-luna named as the replacement for each. These are models that an enormous amount of production code still points at.
The detail most likely to catch people out is one line underneath that table: “Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models.” If you paid to fine-tune a GPT-4o variant, that work stops on 14 January with everything else.
Audio and transcription take a heavier hit than the chat models. Between 6 January and 26 February, tts-1, tts-1-hd, gpt-4o-mini-tts, the legacy realtime and audio families, whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize and the /v1/audio/translations endpoint all go. If you built a transcription pipeline on whisper-1 through the API, 26 February is your date.
Worth separating two things that share a name: the Whisper model weights are open and you can keep running them yourself indefinitely. It is OpenAI’s hosted whisper-1 endpoint that is being withdrawn. For a pipeline that cannot be rewritten by February, self-hosting is a genuine option rather than a consolation.
There is also an escape hatch most developers do not know about: “In some cases, developers may be able to provision dedicated capacity for continued access after a model’s shutdown date.” That is a sales conversation, not a setting, but if a model is genuinely load-bearing for you it is worth having.
Anthropic: the shortest notice, and the most candour
Anthropic commits to “at least 60 days’ notice before model retirement for publicly released models” — a third of OpenAI’s for generally available models. Its lifecycle language is unambiguous about the consequence: “Retired: The model is no longer available for use. Requests to retired models fail.”
The immediate date is 30 November 2026, when claude-sonnet-4-5-20250929 retires with claude-sonnet-5-5 as the recommended replacement. If you are reading this in mid-November, that is weeks away, not months.
Anthropic’s status table carries a subtlety worth understanding. Active models are listed with a “tentative retirement date” phrased as “not sooner than” a given date — claude-opus-4-5-20251101 is “not sooner than November 24, 2026”, and claude-haiku-4-5-20251001 was “not sooner than October 15, 2026”. Those dates are not retirement announcements. They are the earliest point at which one could be made. Several have now passed, which means notice could begin at any time, and sixty days later the model is gone.
One caveat that matters if you are not calling Anthropic directly: “Partner-operated platforms, Claude Models on Google Cloud (Google managed) and Amazon Bedrock, set their own retirement schedules, so a model’s lifecycle status and dates can differ.” Checking Anthropic’s page is not enough if you consume Claude through Bedrock.
Anthropic is also the only one of the three to write down the costs of its own policy — that retirement happens “to ensure capacity for new model releases”, that researchers lose access for comparative work, and that it “introduces safety- and model welfare-related risks”. It says it has committed to long-term preservation of model weights. None of that keeps your production call alive, but it is a more honest framing than the industry norm.
Three ways to find out what you are actually using
Most teams discover a deprecation because something broke, not because they read a changelog. All three platforms give you a way to check first.
OpenAI: query it programmatically
This is the best of the three and almost nobody uses it. “To monitor model shutdowns programmatically, call the List models endpoint (GET /v1/models) and check each model’s optional shutdown_date field. When present, this field contains the shutdown date as a YYYY-MM-DD string, or null if no shutdown date has been announced.”
That is a scheduled job you can write this afternoon: list the models, compare against the model strings in your codebase, and alert when any of them acquires a date.
Anthropic: export your usage
Go to the Usage page in the Claude Console and click Export. The CSV breaks usage down by API key and model, which tells you not only which deprecated models you are calling but which key — and therefore which service — is calling them.
Google: read the changelog, because the API will not tell you
This is the weak spot in the silent re-routing approach. Because nothing errors, there is no signal in your application to react to. The Gemini API release notes are the only place the change is announced. Put them on a reading rota.
One more Google change worth catching
While reading the same changelog we found a default that changed on 26 October 2026 and has nothing to do with retirements: “New projects using the Generate Content API… on the paid tier now store requests by default (store=true), matching the Interactions API. Existing projects keep their current setting.”
If you created a Gemini project after that date and have obligations about where your prompts are retained, this one is yours to check. Google gives two ways out: set store=false per request, or turn off Generate Content API logging in the Logs and Datasets dashboard. Projects created before 26 October are unaffected, which means two projects in the same organisation can behave differently.
What to do this month
- Grep your codebase for model strings. Not just application code — notebooks, cron jobs, serverless functions, CI configuration and anything a colleague wrote last year and left running.
- If you call claude-sonnet-4-5, move before 30 November. This is the only deadline in this article that is inside the month.
- If you call gemini-2.5-pro or 2.5-flash, pin a new model string before 1 December — and budget for the new rates, which are not the old ones.
- Check your fine-tunes. On OpenAI they die with their base model.
- Avoid preview models in production, or accept that two weeks’ notice is the contract.
- Write the shutdown_date check. Ten minutes now, and the next one of these finds you instead of the other way round.
Frequently asked questions
What happens if I call a retired AI model?
It depends entirely on the platform. On Anthropic and OpenAI the request fails — the model is no longer accessible. On Google’s Gemini API the request succeeds, but it is answered by a different model and billed at that model’s rates, with no error raised.
When is GPT-4o being shut down?
14 January 2027, along with GPT-4o mini, GPT-4.1 and GPT-4.1 mini. OpenAI names gpt-6-luna as the recommended replacement for all four. Fine-tuned versions follow the same date as their base model.
How much notice do AI providers give before retiring a model?
OpenAI commits to at least six months for generally available models, three months for specialised and image generation variants, and as little as two weeks for preview models. Anthropic commits to at least 60 days for publicly released models. Google does not publish a general notice period; its recent re-routing notices have given roughly four weeks.
Is whisper-1 going away?
OpenAI’s hosted whisper-1 endpoint shuts down on 26 February 2027, along with gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize. The open Whisper model weights are unaffected and can still be self-hosted.
How do I check which deprecated models my code is using?
On OpenAI, call GET /v1/models and read the shutdown_date field on each model. On Anthropic, export the usage CSV from the Claude Console, which breaks usage down by API key and model. On Google, there is no equivalent signal — you have to read the Gemini API release notes.
Do Bedrock and Vertex follow the same retirement dates?
No. Anthropic states that partner-operated platforms — Amazon Bedrock and Google-managed Claude on Google Cloud — set their own retirement schedules, so lifecycle status and dates can differ from Anthropic’s own page.
What we checked
The Gemini API release notes, OpenAI’s deprecations page and Anthropic’s model deprecations page were read on 10 November 2026, and every date, model name and quoted sentence comes from them. Deprecation schedules change: models are added, dates move, and new announcements appear weekly, so check the three pages above before acting on any date here. We have not tested any of these models, and nothing in this article compares their output.
For what the three platforms’ terms say about your data, see ChatGPT vs Claude vs Gemini. For the cheaper alternatives and which of them you can download and keep, see Grok, DeepSeek and Qwen, and for the case for running models yourself, What self-hosting an AI tool actually costs you.