On October 23, 2026, OpenAI shuts down a long list of older models: GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, o1, o1-pro, o3-mini, o4-mini, an early GPT-4o snapshot, and the fine-tuned versions built on several of them. Any integration that still names one of those models stops getting answers that day. Nothing about it is a surprise; OpenAI announced it in April. It is still the kind of date that catches companies out, because the model name is usually buried in a configuration nobody has opened since the project shipped.
That is one date. Between now and April 2027 there are nine more, across both of the big API vendors. This post lists them, quoted from each vendor’s own deprecation page, and then covers the part that actually takes time: finding out whether any of them is yours.
What “retired” means in practice
A retired model is not slower or worse. It is gone. Requests that name it return an error instead of a response. If the request came from a chatbot, the chatbot stops answering. If it came from an automation that runs overnight, the automation fails, and whether anyone finds out depends on whether that failure goes somewhere a person looks.
Both vendors give notice, and both name a replacement. OpenAI’s page lists a “substitute model” or “recommended replacement” for each model it retires. Anthropic’s says it provides at least 60 days’ notice before retiring a publicly released model. Notice is not the problem. The problem is that the notice goes to whoever owns the API account, and that person is often not the person who knows which of your systems call which model.
The calendar, October 2026 to April 2027
Every date below is taken from OpenAI’s deprecations page or Anthropic’s model deprecations page, as published on October 5, 2026. Vendors add entries, so check the pages themselves before you plan against this list.
Oct 23, 2026 — OpenAI
Stops working: gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini, gpt-image-1, plus fine-tunes of gpt-3.5-turbo, gpt-4, gpt-4.1-nano, o4-mini, babbage-002 and davinci-002.
The vendor names instead: gpt-5.6-sol, gpt-5.6-terra or gpt-5.6-luna, depending on the model; gpt-image-2.5-sunburst or gpt-image-2.5-flare for gpt-image-1.
Oct 31, 2026 — OpenAI
Stops working: Existing evals on the Evals platform become read-only.
Nov 30, 2026 — OpenAI
Stops working: Agent Builder, the v1/prompts API and reusable prompt objects, and the Evals dashboard and API.
The vendor names instead: Agents SDK or ChatGPT Workspace Agents for Agent Builder; prompt content moves into your own code.
Nov 30, 2026 — Anthropic
Stops working: claude-sonnet-4-5-20250929 (Claude Sonnet 4.5).
The vendor names instead: claude-sonnet-5-5.
Dec 1, 2026 — OpenAI
Stops working: Image models gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest.
The vendor names instead: gpt-image-2.5-sunburst or gpt-image-2.5-flare.
Dec 11, 2026 — OpenAI
Stops working: Dated GPT-5 snapshots from 2025 (gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro) and the o3 / o3-pro snapshots.
The vendor names instead: gpt-5.6-sol, terra or luna.
Jan 6, 2027 — OpenAI
Stops working: Text-to-speech: tts-1, tts-1-hd and two gpt-4o-mini-tts snapshots.
The vendor names instead: gpt-realtime-2.1-mini.
Jan 20, 2027 — OpenAI
Stops working: The older realtime and audio families, including gpt-realtime, gpt-4o-realtime, gpt-audio and their mini versions.
The vendor names instead: gpt-realtime-2.1, gpt-realtime-2.1-mini or gpt-audio-1.5.
Feb 26, 2027 — OpenAI
Stops working: Transcription: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize.
The vendor names instead: gpt-live-transcribe or gpt-transcribe.
Apr 1, 2027 — OpenAI
Stops working: gpt-5.1, gpt-5.3-codex, gpt-5.4-nano.
The vendor names instead: gpt-6-sol or gpt-6-luna.
Two things that are deliberately not on that list. First, Anthropic lists many of its active models with a date marked “Not sooner than.” That is the earliest a retirement could happen, not a scheduled one, and treating it as a deadline would have you migrating things that are not going anywhere yet. Second, this covers the two vendors’ own APIs. If you reach a model through a cloud platform or another product, that provider runs its own schedule.
Three entries that are bigger than they look
Fine-tuned models. On October 23, OpenAI removes fine-tuned versions of several older models, and the page names a “recommended replacement base model” for each. A base model is not your fine-tune. If someone trained a model on your data a couple of years ago, the replacement is a project, not a setting: either the newer base model does the job well enough without the training, or the work gets redone.
Voice. The three voice dates land within eight weeks of each other, January 6 to February 26. For text-to-speech, OpenAI tells developers to migrate to a realtime model and points to its Realtime API guide to plan the move. Read that as a migration with testing, not a rename. A phone or voice system usually touches all three: speech in, the model, speech out.
Things that are not models. Agent Builder, reusable prompts and the Evals platform all shut down on November 30. If your team built an agent in a vendor’s visual builder, or kept its prompts in the vendor’s dashboard rather than in your own code, that agent depends on a product with an end date, whatever model it calls.
A replacement is not a drop-in
The vendor’s recommended replacement is a newer, usually better model. It is also a different one. It can format answers differently, follow instructions more literally or less, cost a different amount per request, and take a different amount of time to respond. For a chat window, that may not matter. For an automation that parses the model’s output and writes it into a CRM, it can matter a great deal, and the way you find out is by running your real inputs through it before the date, not after.
That is the same argument we made in what an API integration costs after it ships: the vendor changing the platform is not a failure, it is the deal. The cost is in noticing and in scheduling the work before the date.
What to do this month
- List every place a model name lives. Application code and config files, automation tools such as n8n or Make, chatbot settings, voice platforms, and any vendor product that asked you to pick a model. A name typed into a dropdown two years ago counts. Anthropic’s deprecations page also describes exporting an audit of your API usage from the Usage page in Claude Console, which shows what is actually being called.
- Check it against the list above, then against the vendor pages. Snapshot names matter.
gpt-4o-2024-05-13is on the October list; later GPT-4o snapshots are not. - Find your fine-tunes. If any exist, decide now whether to retrain or retire them, and check whether you still can: OpenAI’s page says active existing customers “will no longer be able to create new fine-tuning jobs” from January 6, 2027, and accounts that have not used a fine-tuned model in the past 60 days already cannot. Retraining may mean a different model or provider, and 18 days is not long for that.
- Test the replacement on real inputs. A dozen examples of what the system actually handles, compared side by side, is usually enough to tell you whether the swap is a setting change or a piece of work.
- Name one owner for vendor notices. Deprecation emails go to the account owner. Someone has to read them and know which systems they touch. That is the job we wrote about in your AI agents need a manager, not just an API key.
None of this needs a program. For most mid-sized companies it is a few hours of looking, and the looking is the part that keeps getting skipped. If you would rather have someone else do the inventory and the testing, that is part of our AI integration work.
Related reading: your AI agents need a manager, not just an API key — why a vendor notice needs an owner on your side. And when your AI vendor goes down — the unplanned version of the same dependency. And which model is best — the question a replacement reopens. The service page for this work is AI integration.