OpenAI o3 Retired From ChatGPT: What Changes Now

OpenAI o3 retired from ChatGPT on 26 August 2026, closing a 90-day sunset that started on 28 May. If you reached for o3 in the model menu when you wanted slower, more careful reasoning, that entry is gone from the web and mobile apps. Your API traffic is untouched. The o3 endpoint keeps serving until 11 December 2026, per OpenAI's own deprecations page. Two surfaces, two clocks, roughly three and a half months apart.

Why o3 retired from ChatGPT does not mean o3 is gone

The confusion is structural. Most people meet a model through a product and assume the product is the model. When the name disappears from a dropdown, it reads as a shutdown. It usually is not.

OpenAI runs consumer retirement and API retirement on separate schedules because they answer to different pressures. The ChatGPT model list is a user experience problem: too many entries and nobody picks well. The API is a contract problem: code depends on a string, and breaking that string breaks production. So the picker gets pruned aggressively and the endpoint gets a long, published runway.

The practical consequence is that the same word, retired, means "you cannot click it any more" in one place and "your integration returns a 404 in December" in the other.

The four o3 clocks

Every date below comes from OpenAI's deprecations page or from Forkast's 26 August report on the retirement.

The o3-pro carve-out is the one people miss. If your team pays for Pro or sits on an Enterprise plan, the heaviest o3 variant is still there. What went away is the middle tier.

The same day took the Assistants API too

26 August was a double date. The Assistants API also shut down, exactly one year after it was announced for retirement in August 2025, with the Responses API and Conversations API named as replacements. That one had a full twelve months of notice and still caught teams mid-migration, which tells you something about how much attention deprecation notices get when the endpoint keeps working.

If you were running both, you had a hard break and a soft break landing on the same morning. The hard break is the one that pages you.

What to check this week

1. Grep your codebase for the literal string o3. Model names get pasted into config files, environment variables, prompt templates and analytics dashboards, and only some of those are in the repo you are thinking of.

2. Check whether anything routes to o3 by default when a preferred model is unavailable. Fallback chains are where dead model names survive longest, because they only execute when something else has already gone wrong.

3. Run your evals against gpt-5.6-sol before December rather than after. A replacement is not a drop-in until you have measured it on your own tasks, and reasoning models in particular shift behaviour on edge cases in ways an aggregate benchmark hides.

4. Tell your users if the change is visible to them. If a customer-facing feature has been quietly leaning on o3-mini, October is closer than it looks.

Why this keeps happening

Model retirement is now a recurring operational cost rather than an occasional event. The o3 family alone generated five distinct dates across two surfaces in seven months. Anyone building on top of hosted models is running a small, permanent migration programme whether they have named it or not.

The habit worth forming is boring: keep the model identifier in one place, keep a dated list of what you depend on, and check the vendor deprecation pages on a schedule instead of when something breaks. We wrote up the general version of that in what to do when an AI model gets deprecated, and the migration mechanics in moving from one AI model to another.

For context on how this compares with the last two retirements, see our coverage of the gpt-5.2-chat-latest retirement and the Assistants API shutdown. If you want a lighter-weight way to keep track of all of it, our guide to keeping up with AI news covers what is worth watching and what is noise.