Within two weeks this autumn, two of the three big model providers put end dates on models that a lot of business chatbots still run on. On 30 September 2026 Anthropic notified developers that Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires on 30 November 2026. On 14 October OpenAI told developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini will be removed from its API on 14 January 2027. A retired model does not give you a slightly worse answer. Anthropic's page puts it plainly: "Requests to models past the retirement date will fail."
If your support bot, internal assistant or document pipeline was built in 2024 or 2025 and nobody has opened its model setting since, something in the calendar below probably applies to you. Every date comes from the providers' own deprecation pages, checked on 1 November 2026.
The shutdown calendar, November 2026 to May 2027
| Date | Provider | Model or feature | Named replacement |
|---|---|---|---|
| 17 Nov 2026 (earliest) | Google Gemini API | gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts, gemini-3.1-flash-tts-preview | gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts |
| 17 Nov 2026 (earliest) | Google Gemini API | gemini-3.1-flash-live-preview, gemini-2.5-flash-native-audio-preview-12-2025 | gemini-3.8-live |
| 30 Nov 2026 | Anthropic | claude-sonnet-4-5-20250929 | claude-sonnet-5-5 |
| 30 Nov 2026 | OpenAI | Reusable prompt objects and the v1/prompts API, Agent Builder, the Evals dashboard and API | Prompts in your own code, the Agents SDK or ChatGPT Workspace Agents, a Promptfoo migration path |
| 11 Dec 2026 | OpenAI | gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, o3-pro | gpt-5.6-sol, gpt-5.6-terra or gpt-5.6-luna |
| 6 Jan 2027 | OpenAI | tts-1, tts-1-hd, gpt-4o-mini-tts | gpt-realtime-2.1-mini |
| 14 Jan 2027 | OpenAI | gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, and fine-tunes built on them | gpt-6-luna |
| 20 Jan 2027 | OpenAI | gpt-realtime, gpt-audio, gpt-4o-realtime, gpt-4o-audio and their mini versions | gpt-realtime-2.1, gpt-realtime-2.1-mini or gpt-audio-1.5 |
| 26 Feb 2027 | OpenAI | whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, /v1/audio/translations | gpt-live-transcribe or gpt-transcribe, v1/realtime/translations |
| 1 Apr 2027 | OpenAI | gpt-5.1, gpt-5.4-nano, gpt-5.3-codex | gpt-6-sol or gpt-6-luna |
| 7 May 2027 (earliest) | Google Gemini API | gemini-3.1-flash-lite | gemini-3.5-flash-lite |
How firm are these dates?
It depends on who wrote them. Google says the dates in its table "indicate the earliest possible dates on which a model might be retired" and that the exact date will come with advance notice. Treat 17 November as the day your voice feature could stop, not the day it will. Google also says its Gemini 2.5 models "are not deprecated and will continue to be served until further notice", but access is limited to users who have used them before. A new project cannot start on them.
OpenAI's dates are fixed shutdowns. The 1 October batch came "with six months' notice". The GPT-4 series got three months, from 14 October to 14 January.
Anthropic promises "at least 60 days' notice before model retirement" for publicly released models, and lists active models with a "not sooner than" date. Two of those dates have arrived or nearly have: claude-haiku-4-5-20251001 shows "Not sooner than October 15, 2026" and claude-opus-4-5-20251101 shows "Not sooner than November 24, 2026", on the same page. Neither is deprecated yet. With the 60-day minimum, a notice sent today could not retire either before the end of December. If a cheap Haiku model sits behind your FAQ bot, expect that email and plan for it now.
The failures that do not show up as errors
A hard shutdown is the easy case. Your logs fill with errors and somebody notices. These four problems are quieter.
Fine-tuned models go with their base
OpenAI states that "fine-tuned versions of these models follow the same shutdown schedule as their underlying base models". If you trained a GPT-4o mini on your own tickets to get the tone and the product names right, that model ends on 14 January 2027 with the rest. You need the training data, a fine-tuning run on whatever replacement you pick, and time to test it. Check first that your chosen replacement can be fine-tuned at all.
Model names that point somewhere new
Some IDs keep working while the model behind them changes. Google's table says the gemini-3.5-flash endpoint "is not shutting down and automatically routes API requests to gemini-3.6-flash", and gemini-3.7-flash now routes to gemini-3.8-flash. Nothing breaks. But the model answering your customers is not the one you tested, and changes in answer length, refusals or formatting arrive without a deploy on your side. If a Gemini bot's behaviour shifted recently and nobody changed any code, check this first.
A dependency bump that changes the API version
Anthropic released API version 2026-10-21. Its migration guide says that if you use an SDK, "your requests use the API version that was current when your SDK's major version was released". In the new version, "a request that sends temperature, top_p, or top_k returns a 400 error, even at the parameter's default value". Plenty of chatbot code sets temperature out of habit. An automated pull request that moves the SDK to its next major version can therefore break every call, even though the model never changed.
There is a second effect. On models that still accept other values, such as Claude Sonnet 4.6, a lowered temperature that kept answers consistent simply stops applying, and the guide says "No other setting replaces it." You can keep such requests on 2023-06-01 for now. Decide that on purpose rather than finding out from your users.
Cloud marketplaces keep their own calendar
Anthropic's dates apply to its own platforms. Amazon Bedrock and Google-managed Claude on Google Cloud "set their own retirement schedules", so the same model can have a different status and date there. If your company buys Claude through an AWS account, read the Bedrock model table. The Anthropic page alone does not tell you your date.
Voice bots get hit more than once
Text chat sees one or two dates. A voice assistant on OpenAI can see three: text-to-speech models end on 6 January 2027, the older realtime and audio models on 20 January 2027, and whisper-1 plus the 4o transcription models on 26 February 2027. On Gemini, the preview TTS and live models carry the 17 November 2026 earliest date. The TTS replacement OpenAI names is a realtime model, so this is a change of interface, not a one-line rename. Teams doing call summaries or Arabic voice menus should start with this part, because a new voice model may also sound different, and callers notice that faster than a change in wording.
Checklist: what to change, and by when
- List every model ID you call, per API key. This week. In Anthropic's Console, the Usage page has an Export button that gives a CSV "broken down by API key and model". Then search code, config files, low-code workflows and scheduled scripts. Ask each SaaS vendor whose AI feature you rely on which model it uses and when they will move.
- Move the model ID into configuration. One setting per use case, holding a dated snapshot name. Swapping models should be a config change you can roll back in minutes, not a code release.
- Build a test set before you switch. Take real questions from your logs, in every language your bot serves, with the answers you would accept. Include questions it must refuse. Our article on assistants that cite or refuse explains why refusals belong in the set.
- Run the replacement next to the old model. Compare correctness, refusals, answer length, latency and cost per conversation. Look up the replacement's price on the provider's pricing page; don't assume it matches.
- Remove
temperature,top_pandtop_kfrom Claude requests, or pin the SDK until you have tested API version2026-10-21. - Before 30 November: move off Claude Sonnet 4.5, and copy any prompts stored as OpenAI prompt objects into your own code.
- Before 11 December: replace
gpt-5,gpt-5-mini,gpt-5-nano,o3and their pro versions. - Before 6 January: start retraining fine-tunes and migrating voice, since GPT-4o models, TTS, realtime and transcription all end between 6 January and 26 February.
- Give the calendar an owner. Anthropic says impacted customers "will always be notified by email". If the address on the account belongs to a former employee or an unread billing inbox, that notice reaches nobody.
Our view
A model ID is now a dependency with an expiry date, like a TLS certificate or a database version. The teams that handle this calmly are the ones that already have the test set and the config switch, so a retirement notice becomes a one-day job. Teams without them get the same notice and spend weeks arguing over whether the new model is "worse". If you are weighing a hosted model against running one inside your own network, our comparison of on-premise and cloud deployment for an AI assistant covers the trade-off, including who controls the upgrade schedule. And after any switch, measure whether the bot still closes requests, not only whether it replies: the difference is explained in our piece on ticket deflection versus resolution.
Frequently asked questions
What happens if my chatbot still calls a retired model?
The request fails. Anthropic's deprecations page states that requests to models past the retirement date will fail, and OpenAI describes its dates as removal from the API. The bot will return an error or whatever fallback message your code shows, so customers see a broken assistant rather than a weaker one.
When does Claude Sonnet 4.5 retire, and what replaces it?
Anthropic lists 30 November 2026 as the retirement date for claude-sonnet-4-5-20250929 on its own platforms, with claude-sonnet-5-5 as the recommended replacement. If you use Claude through Amazon Bedrock or Google-managed Google Cloud, check that platform's model table, because they set their own schedules.
Are GPT-4o and GPT-4.1 being removed from the OpenAI API?
Yes. OpenAI notified developers on 14 October 2026 that gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini will be removed on 14 January 2027, with gpt-6-luna as the recommended replacement. Fine-tuned versions of these models shut down on the same date.
Will the 17 November date really switch off Gemini voice models?
Possibly, but not necessarily on that day. Google describes the dates on its deprecations page as the earliest possible shutdown dates and says the exact date will be communicated with advance notice. Plan as if 17 November is real for the preview TTS and live models.
Why did our Claude integration start returning 400 errors after an SDK update?
A new SDK major version can move your requests to API version 2026-10-21, in which sending temperature, top_p or top_k returns a 400 error even at default values. Remove those parameters, or keep the request on API version 2023-06-01 until you have tested the change.
Sources
- Anthropic: Model deprecations
- Anthropic: Migrating to API version 2026-10-21
- OpenAI: Deprecations
- Google AI for Developers: Gemini deprecations
All dates and model names checked against the providers' pages on 1 November 2026. Providers can change schedules; their pages are the final word.


