AI Model Update FAQs: Deprecations, Rollbacks and Why Answers Change Overnight
Deprecated, shut down or retired? OpenAI's 6-month, 3-month and 2-week notice clocks, Anthropic's 60-day rule, Gemini reroutes that change your model and bill, rollbacks — and why answers change under the same name.
Your AI gave a different answer this morning to the same question it handled fine last night — and in most cases nothing broke: a model was updated, rerouted or quietly fixed behind the same name.
Key points
- Deprecated means a retirement date is set and the model still works; shutdown or retired means requests fail.
- Warning time varies: OpenAI gives 6 months for general models, Anthropic 60 days, previews about 2 weeks.
- Answers can change under an unchanged name through rerouting, a latest alias swap, or serving changes.
Deprecation, shutdown and retirement do not mean the same thing
Deprecated is a warning with a date; shutdown and retired are the end.
The words sound interchangeable, and that is why they confuse people. OpenAI sorts every model into three states: Active, where no retirement has been announced; Deprecated, where retirement has been announced but you can still use the model until its shutdown date; and Shutdown, where the model no longer accepts requests.
Google defines the same arc plainly: a deprecation announces that support will end and the model will be shut down soon, and once shut down it is completely turned off. Our pillar on how AI model updates work explains where these endings fit.
OpenAI: three states and three different warning clocks
General models get 6 months, specialised variants 3 months, previews about 2 weeks.
OpenAI says impacted customers are always notified by email and in its documentation, and that, unless safety or compliance concerns require a faster timeline, it provides minimum notice periods. Generally available models get at least 6 months. Specialised variants and image generation models get at least 3 months, with chat variants such as gpt-5.1-chat-latest as an example.
Preview models may be retired with much shorter notice, such as 2 weeks, and OpenAI does not recommend them for business-critical workloads. OpenAI also lets you check programmatically: the List models endpoint, GET /v1/models, exposes each model’s optional shutdown_date field. Its page lists gpt-5 shutting down Dec 11, 2026 and GPT-4o on Jan 14, 2027.
Google: some models end, others quietly reroute to a new one
A shutdown turns the endpoint off; a reroute runs your code on a different model.
Google adds a twist the word shutdown does not capture: rerouting. Its deprecations page says gemini-3.7-flash is no longer available in AI Studio, yet its endpoint is not shutting down and automatically routes requests to gemini-3.8-flash. The same applies to gemini-3.5-flash, which routes requests to gemini-3.6-flash.
The bigger reroutes are dated. Effective December 1, 2026, all requests to gemini-2.5-pro will automatically route to gemini-3.8-flash, and all requests to gemini-2.5-flash will automatically route to gemini-3.5-flash-lite. Nothing errors — your request succeeds, but a different model answers it. Google warns its listed shutdown dates are the earliest possible dates. To see which name you are actually on, follow how to tell which AI model version you are using.
Claude: four stages, at least 60 days, and a platform catch
Active, Legacy, Deprecated, Retired — and the dates cover Anthropic-operated platforms only.
Anthropic uses four stages: Active means fully supported and recommended; Legacy means the model will no longer receive updates and might be deprecated later; Deprecated means still functional but no longer recommended, with a replacement and a retirement date; and Retired means no longer available, with requests failing.
Its notice promise is one flat number: at least 60 days’ notice before retirement for publicly released models. For example, Claude Sonnet 4.5, claude-sonnet-4-5-20250929, was deprecated on September 30, 2026 with a tentative retirement date of November 30, 2026 and claude-sonnet-5-5 as its replacement. The catch is scope: the dates apply to Anthropic-operated platforms; partner platforms such as Amazon Bedrock set their own schedules.
Why answers change even when the model name does not
Fixed weights can still sit behind a moving router, new safety classifiers or a swapped alias.
Sometimes nothing was deprecated, and the answers still moved. Anthropic explains why: each Claude model ID identifies a pinned version, the underlying model remains constant for that ID’s lifetime, and Anthropic does not update the weights of an existing ID — yet the serving infrastructure around it, including the request router, safety classifiers and sampling logic, can change and produce minor differences in behaviour.
Google’s labels create the same effect by design. A stable name points to a specific model that usually does not change, but a latest alias such as gemini-flash-latest is hot-swapped with every new release, with a 2-week email notice. OpenAI’s October 29, 2026 release notes say GPT-6 in Chat was updated for Free and Go users — more concise responses, less flattering language — while the main GPT-6 model on Plus and Pro was unchanged.
Rollbacks and quiet fixes: when an update is undone or corrected
Providers do reverse changes — and the correction is logged in the release notes.
Not every change sticks. OpenAI’s release notes record a clear rollback case: on January 10, 2026 it lowered Standard and Light thinking time, the Extended setting was unintentionally lowered in that update, and on February 4, 2026 it restored Extended to its prior level, correcting the inadvertent reduction.
The same notes show quieter fixes where nothing is announced as broken, such as the GPT-5.5 Instant update of May 28, 2026. When answers shift, look first at the provider’s own release notes — the change is usually written down there.
What this means
The model is rented, not owned — so the calm move is to note the name and expect an ending.
It is a strange, helpless feeling the first time it happens to you. Yesterday the assistant drafted replies in exactly your tone. Today the same prompt comes back shorter, blunter, or oddly eager to please, and you catch yourself wondering whether you imagined the old version. You did not. Something above your conversation changed — a setting, a route, a whole model — and nobody knocked on your door first.
Once you see the machinery, the feeling changes. A deprecation notice is not the product failing you; it is the one part of this system that actually respects your time, because it gives you a date, a replacement and a window to adjust. The genuinely awkward changes are the silent ones: a Gemini name rerouted to a newer model, a latest alias swapped underneath your app, a behaviour tune that reaches free accounts one week and paid accounts another. None of those send an error. That is also why two colleagues on different plans can compare notes, disagree completely about whether the assistant got worse this week, and both be honestly describing what they saw on their own screens.
So the practical habits are small and unglamorous. Write down the full model name behind any workflow you rely on, the way you would note which version of a recipe actually worked. If answers drift, check that name first, then the deprecations page and release notes, before rewriting prompts in a panic. Treat preview and latest labels as a promise that change is coming soon, and give important work to stable, pinned names instead. And when a retirement date appears, believe it: the morning a retired endpoint starts failing is a miserable time to discover your favourite tool had an expiry date all along.
Bottom line
Read the lifecycle word and the date.
Deprecated means plan; shutdown and retired mean stop; an unchanged name guarantees nothing about unchanged behaviour. Note the exact model behind work you care about, favour stable and pinned names, and check the official pages first when answers change overnight.
Frequently Asked Questions
Billing, permanence and ChatGPT-versus-API questions not covered above.
Will you be charged differently if your Gemini model is rerouted?
Yes, you can be charged at the replacement model’s rates after a reroute. Google says gemini-2.5-pro requests will route to gemini-3.8-flash from December 1, 2026 and be billed at Gemini 3.8 Flash rates, while gemini-2.5-flash requests will be billed at gemini-3.5-flash-lite rates.
Can you keep using a deprecated model forever if you never update anything?
No, you cannot keep a deprecated model forever by leaving your settings untouched. Deprecated means the model still works only until its shutdown or retirement date. After that date, OpenAI shutdown models accept no requests, Anthropic retired models fail, and Google endpoints are unavailable.
Does retiring a model in ChatGPT also remove it from the API?
No, a ChatGPT retirement does not automatically remove the same model from the API. OpenAI’s release notes say retiring GPT-5.5 from ChatGPT, ChatGPT Work and Codex on October 14, 2026 does not affect API access, and its February 2026 GPT-4o retirement likewise brought no API changes.
{“@context”:”https://schema.org”,”@type”:”FAQPage”,”mainEntity”:[{“@type”:”Question”,”name”:”Will you be charged differently if your Gemini model is rerouted?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”Yes, you can be charged at the replacement model’s rates after a reroute. Google says gemini-2.5-pro requests will route to gemini-3.8-flash from December 1, 2026 and be billed at Gemini 3.8 Flash rates, while gemini-2.5-flash requests will be billed at gemini-3.5-flash-lite rates.”}},{“@type”:”Question”,”name”:”Can you keep using a deprecated model forever if you never update anything?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”No, you cannot keep a deprecated model forever by leaving your settings untouched. Deprecated means the model still works only until its shutdown or retirement date. After that date, OpenAI shutdown models accept no requests, Anthropic retired models fail, and Google endpoints are unavailable.”}},{“@type”:”Question”,”name”:”Does retiring a model in ChatGPT also remove it from the API?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”No, a ChatGPT retirement does not automatically remove the same model from the API. OpenAI’s release notes say retiring GPT-5.5 from ChatGPT, ChatGPT Work and Codex on October 14, 2026 does not affect API access, and its February 2026 GPT-4o retirement likewise brought no API changes.”}}]}
Related Guides
Keep exploring how models update and end.
- AI Model Updates Explained: How ChatGPT, Gemini and Claude Get New Versions
- How to Tell Which AI Model Version You Are Using (and Why It Sometimes Changes)
- AI Model Names Explained: What GPT, Gemini, Claude and Llama Version Numbers Actually Mean
- Mistakes to Avoid When a New AI Model Launches
- How to Switch AI Models Without Losing Your Chats
Sources
The official pages behind every claim above.
