Models do not last, and both major vendors publish exactly how long you get. OpenAI commits to at least six months for a generally available model. Anthropic commits to at least sixty days. Neither number is a guess — both are written down, and the second is about a third of the first.
The notice period is a published number. Most people never read it.
- OpenAI: at least 6 months for generally available models.OpenAI, Deprecations, read at source 16 Sep 2026: “Generally available models: At least 6 months.”
- Anthropic: at least 60 days for publicly released models.Anthropic, Model deprecations, read at source 16 Sep 2026: Anthropic provides “at least 60 days’ notice before model retirement for publicly released models.”
- Preview models get two weeks on OpenAI, and it says not to build on them.
- Retired means broken. Both vendors say requests simply fail.
- Anthropic names the downside out loud, which is rare.
How much warning you actually get
| Kind of model | OpenAI | Anthropic |
|---|---|---|
| Generally available | at least 6 months | at least 60 days |
| Specialised variant | at least 3 months | not stated separately |
| Preview | as little as 2 weeks | not stated separately |
OpenAI, Deprecations · Anthropic, Model deprecations. Both read at source 16 Sep 2026.
Both carry the same escape clause. OpenAI: “If safety or compliance concerns require us to retire a model sooner, we will provide as much notice as reasonably possible.” The floor is a commitment in ordinary conditions, not a guarantee.
Sixty days is about two months. Six months is two quarters you can plan around. If you are choosing between vendors for something long-lived, that difference belongs in the decision alongside the price.
What “retired” actually means
Not slower, not deprioritised. Gone. Anthropic’s three states are worth knowing because they are the warning ladder:
- Legacy — “The model will no longer receive updates and may be deprecated in the future.”
- Deprecated — “The model is still functional but no longer recommended.” A retirement date is assigned at this point.
- Retired — “The model is no longer available for use. Requests to retired models will fail.”
OpenAI uses different words for the same end state, treating “sunset” and “shut down” as interchangeable, “to mean a model or endpoint is no longer accessible.”
How fast it has actually happened
The published history is more useful than the policy, because it shows the real gap between the two announcements.
| Model | Deprecated | Retired |
|---|---|---|
| Claude Opus 4.1 | 5 Jun 2026 | 5 Aug 2026 |
| Claude Sonnet 4 | 14 Apr 2026 | 15 Jun 2026 |
| Claude Opus 4 | 14 Apr 2026 | 15 Jun 2026 |
Anthropic, Model deprecations, read at source 16 Sep 2026, from the model status table.
Two months, in each case. A model announced in spring was unusable by summer. If your product pinned a dated snapshot, that is the window you had.
OpenAI’s own list has several shutdowns inside the next three months:
| Model | Shuts down |
|---|---|
gpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-1106 | 28 Sep 2026 |
gpt-5.4-cyber | 1 Oct 2026 |
gpt-5-nano-2025-08-07 | 11 Dec 2026 |
OpenAI, Deprecations, read at source 22 Sep 2026: the four legacy models each “2026-09-28” with recommended replacement gpt-5.6-terra; “The gpt-5.4-cyber model is deprecated and will be removed from the API on October 1, 2026. Migrate to gpt-5.6-cyber before the shutdown date.”; gpt-5-nano-2025-08-07 “Dec 11, 2026”, replacement gpt-5.6-luna.
The spread is the point. The four legacy models were announced on 26 September 2025, a year ahead. The nano snapshot was announced on 11 June 2026, six months ahead, which is exactly the floor. gpt-5.4-cyber was announced on 11 September 2026 for removal on 1 October — twenty days. The deprecations page does not say which notice class that model falls under.OpenAI, Deprecations, read at source 22 Sep 2026: section headings “2025-09-26: Legacy GPT model snapshots”, “2026-06-11: GPT-5 and o3 model deprecations” and “2026-09-11: GPT-5.4-Cyber”.Reasoning, 22 Sep 2026 — the gaps are day counts between each heading date and its shutdown date.
The part a vendor did not have to write
Anthropic’s page includes a section on the costs of its own policy, which is unusual enough to quote. It states plainly why it retires models at all — “to ensure capacity for new model releases” — and lists what that costs, including that “Model retirement introduces safety- and model welfare-related risks”.
That is a vendor naming a downside of its own decision. It does not make the retirement less disruptive, but it is the opposite of the usual framing, and worth noticing.
What to do about it
- Never pin a
previewmodel in production. OpenAI says so itself: it does not recommend them “for business-critical production workloads unless you can migrate on short notice”. - Check what you are actually calling. Anthropic offers an audit of your API usage specifically to find deprecated models still in use.
- Test the replacement before the date, not after. Both vendors say the same thing, and it is the step people skip.
- Watch the cheap model. It is often the oldest — Claude Haiku 4.5 has the earliest retirement floor of the current Claude family, and the snapshot behind OpenAI’s
gpt-5-nanoshuts down on 11 December 2026.OpenAI, GPT-5 nano model page, read at source 22 Sep 2026: “Default snapshot:gpt-5-nano-2025-08-07”.
The live example this month is Sora and the Videos API, which shuts down on 24 September with no replacement named. Four days later, on 28 September, the four legacy models in the table above go too.
What this page could not verify
- Google and xAI’s notice policies. Not found as a published commitment; only OpenAI and Anthropic state a minimum.
- Whether the minimums are always honoured. The history here is the vendors’ own record of it.
- What happens to fine-tuned models in every case. OpenAI addresses it per deprecation rather than as one rule.
OpenAI, Deprecations · Anthropic, Model deprecations. Both read at source on 16 September 2026.
Notice periods are policy, not law, and policies change.