October 2026 AI model retirements: what changes on OpenAI, Claude, Gemini and Mistral

If your product calls an AI model, October and November are a busy stretch. Between now and the end of November, OpenAI and Anthropic retire models and features, Google has already changed who gets which Gemini model in its app, and Mistral has opened a new flagship in preview. None of it is dramatic on its own, but a hard-coded model name that stops working is exactly the kind of bug that shows up at 2 AM.

Here is what changes, by date, and what to check in your own code.

The dates at a glance

Date	Vendor	What happens
9 Oct (already live)	Google	Gemini app access changes for free personal accounts
14 Oct	OpenAI	GPT-5.5 leaves ChatGPT, ChatGPT Work and Codex (not the API)
23 Oct	OpenAI	A batch of legacy API snapshots shuts down
31 Oct	OpenAI	Evals platform goes read-only
30 Nov	OpenAI	Evals, Agent Builder and reusable prompts shut down
30 Nov	Anthropic	claude-sonnet-4-5-20250929 retires

OpenAI: two separate retirements, easy to mix up

In ChatGPT and Codex. OpenAI says GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on 14 October, across all plans including Business, Enterprise and Edu. Paid users are pointed to GPT-6 Sol, and Free and Go users to GPT-6 Luna. The same page is explicit that this does not apply to the OpenAI API, so if you only call GPT-5.5 through the API, nothing breaks on the 14th.

One thing worth noting: some third-party coverage named a different replacement model. When in doubt, read the vendor’s own page.

If your team uses Codex while signed in with ChatGPT, check anything that names GPT-5.5 explicitly: workspace defaults, saved settings and scheduled tasks.

In the API. The bigger cleanup is on OpenAI’s deprecations page. On 23 October, a long list of legacy snapshots shuts down, including gpt-4-turbo, gpt-4-0613, gpt-4.1-nano, gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini, gpt-image-1, and a set of fine-tuned (ft-*) models. The suggested replacements are mostly the GPT-5.6 family (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna). For gpt-image-1, they are gpt-image-2.5-sunburst or gpt-image-2.5-flare.

Fine-tuned models deserve special attention. You can’t move a fine-tune to a new base by changing a string. If something you rely on is a fine-tuned model on that list, you need a plan this week, not on the 22nd.

The same page lists platform features going away:

• Evals becomes read-only on 31 October and shuts down on 30 November, with a migration path to Promptfoo.
• Agent Builder shuts down on 30 November. The Agents SDK or ChatGPT Workspace Agents are the suggested alternatives.
• Reusable prompts (v1/prompts) also end on 30 November. The advice is to move that prompt content into your own application code.

Looking further out, notices published on 1 October retire the older text-to-speech models (tts-1, tts-1-hd and two gpt-4o-mini-tts snapshots) on 6 January 2027, and gpt-5.1, gpt-5.3-codex and gpt-5.4-nano on 1 April 2027.

Anthropic: one real deadline, two “not before” dates

On Anthropic’s model deprecations page, claude-sonnet-4-5-20250929 was deprecated on 30 September and retires on 30 November 2026. The recommended replacement is claude-sonnet-5-5.

Two other entries are easy to misread:

• claude-haiku-4-5-20251001 is still active. Its tentative retirement date is “not sooner than 15 October 2026”, which is the earliest possible date, not a scheduled shutdown.
• claude-opus-4-5-20251101 is also active, with “not sooner than 24 November 2026”.

Anthropic’s stated policy is at least 60 days’ notice before retiring a publicly released model, so neither of these can disappear without warning. Still, they’re worth putting on your list.

The same page has a quieter change that can break working code. On Claude 4.7 and later models, temperature, top_p and top_k are deprecated, and setting them to non-default values returns a 400 error. If you’re upgrading from Sonnet 4.5 to Sonnet 5.5 and your request still sets temperature: 0.2, remove it and steer behaviour through the prompt instead.

Google: a consumer change, not an API one

Google’s change is about the Gemini app for personal accounts, not the Gemini API. From 9 October, users without a Google AI plan mostly get Flash-Lite. With the default Auto setting, harder prompts may still be routed to Flash or Pro. Google AI Pro and Ultra subscribers keep Flash-Lite, Flash and Pro. AI Plus subscribers are told to expect an email with their own date.

This matters if your support team, sales team or customers use the free Gemini app for work and have noticed the answers changing. The fix there is a plan decision, not a code change.

Mistral: a new option to test

On 6 October, Mistral released a public preview of Mistral Large 4 through its Studio API. It is priced at $1.36 per million input tokens and $4.18 per million output tokens, and Mistral says open weights will follow by the end of the month. If you were going to re-test prompts anyway because of the retirements above, it’s a reasonable model to add to that comparison, especially if you need the option to self-host later.

A 20-minute checklist

1. Search your code for model names. grep for gpt-4, gpt-4o-2024, o1, o3-mini, o4-mini, gpt-image-1, ft: and claude-sonnet-4-5. Also check environment variables and config tables, not just source files.
2. List your fine-tunes and their base models. Anything on a retiring base needs a re-training or re-prompting plan before 23 October.
3. Check sampling parameters before moving to Claude 4.7 or later. Remove temperature, top_p and top_k if they’re set to non-default values.
4. Move anything that depends on OpenAI Evals, Agent Builder or stored prompts before 31 October (Evals) or 30 November (the others).
5. Re-run your own test prompts on the replacement model. Replacements are usually better, but “better” can still mean a different output format, longer answers or a new refusal, which breaks parsers.
6. Add a reminder for 30 November and for 1 April 2027.

None of these dates should surprise anyone who reads the vendors’ deprecation pages, but very few teams do. Bookmark the vendor pages linked above and check them once a month.