How long does an AI model live?

Every app built on a model API is built on something with an expiry date. OpenAI scheduled the API name gpt-4 to stop answering on 23 October 2026, 1,319 days after GPT-4 launched. Most models don't get anywhere near that long. Providers ship a new generation every few months, and each one sends an older model toward a deprecation notice, a shutdown date and, for whoever still calls it, a migration.

We put every public model retirement we could date into one dataset: 341 model ids from OpenAI, Anthropic, Google (the Gemini API and Vertex AI), Mistral, xAI, Cohere and Amazon Bedrock, read from the providers' own deprecation pages, changelogs and archived copies of them, up to 23 October 2026. Our headline numbers come from the 204 generally available models with a known launch date on six first-party APIs. The answers to three questions are below: how long a model lives, how much notice you get, and whether it's getting shorter. The methodology explains how we counted.

How long a model lives#

A model's life here runs from the day its id could first be called on the provider's own API to the day requests stopped being served by it. Many models in the dataset are still running, so a plain average of retired models would read too short: the recent ones that are still alive would be missing. We used a Kaplan–Meier survival estimate instead, the standard method when some subjects haven't reached the end yet. It counts every model for as long as we've watched it.

The curve is flat for the first six months: almost nothing is retired before it's half a year old. The drop starts around the first birthday. That is when Anthropic's "not sooner than" dates and Google's 12-month floor on Vertex AI expire, and when a model's successor has usually shipped. By two years, four in five models are gone.

The providers differ a lot. OpenAI keeps models running longest. Its 2022 embedding and GPT-3 models ran for close to two years, and the GPT-3 base models it opened the API with in 2020 lasted 1,302 days. Mistral's models had the shortest lives of the labs with many retirements: half were gone within a year. xAI's five retired model ids lived shorter still, but that rests on one retirement event.

The pooled 555 days holds up within about a hundred days when we change how we count. Treating the 10 OpenAI ids in the set that were scheduled to shut down on 23 October as still alive gives 562 days. Counting each release batch once instead of each model id gives 487. Dropping Mistral, whose launch dates are approximate, gives 660.

Is it getting shorter#

Yes. Grouped by the year a model launched, the median lifetime has fallen by almost half in three years.

The clearest sign is how soon a model gets its deprecation notice. Among models launched in 2022, one in 27 had been deprecated within a year. For 2025's launches, it was two in three. Among models that have been deprecated, the time from launch to the announcement fell from a median of 527 days for 2022's models to 295 days for 2025's. Retirement within the first year didn't happen in our data until 2024: no 2022 or 2023 model was switched off before its first birthday, but 20% of 2024's and 37% of 2025's were.

Claude shows the trend most cleanly because Anthropic retires models in tidy generations. Each Claude 3 model ran for 500 to 770 days. The Claude 4 generation got about a year: Sonnet 4 and Opus 4 were retired 389 days after launch, and Opus 4.1 exactly 365 days after its own. None has lived less than 360 days. Each Claude model now lists a "not sooner than" retirement date about a year after its release, and the most recent retirements land close to it.

More models also means more retirements. Across the six APIs, 35 model ids were retired in 2024, 44 in 2025 and 61 in 2026 to 23 October, 10 of them scheduled for that day. New generations arrive fast: in our state of AI coding agents we found frontier models' METR time horizon doubling about every four months, and each step makes the previous model easier to retire. Anthropic gives a reason for retiring models at all: the cost of serving them "scales roughly linearly with the number of models we serve."

How much notice you get#

Notice is the time between a provider announcing a retirement and switching the model off. We counted it per announcement: when OpenAI retired 28 models on one date with one notice, that's one decision, not 28. The median across 47 announcements is 114 days, a little under four months. Counted per model id it's 181 days, because OpenAI retires models in large batches with six months' notice.

The notice has been getting shorter. In 2023, the median announcement gave 312 days. In 2026 it gave 95. OpenAI's median per model has stayed near six months, so most of the decline comes from the other providers. Their median fell from 182 days in 2024 to 69 in 2025 and 62 in 2026.

Anthropic is the clearest case. Until August 2025 its deprecations page promised "at least 6 months" of notice; the 7 August capture still says so. By the 10 August capture it said 60 days, and three days later Anthropic deprecated both Claude 3.5 Sonnet models with 76 days' notice. We found no announcement of the change outside the docs. Anthropic has kept the new promise: all 20 of its GA deprecations gave at least 60 days. It now gives almost exactly the minimum.

OpenAI moved the other way, at least on paper. Its deprecations page had no written minimum until June 2026: the 29 May capture has none, and the 3 June capture says GA models get "at least 6 months", specialized variants three months, and previews possibly as little as two weeks. On 14 October 2026 OpenAI notified developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots would be removed on 14 January 2027: 92 days for GA models. That falls short of the minimum the same page states, which allows exceptions only for "safety or compliance concerns". The entry doesn't cite either.

Short notice has a history. In March 2023 OpenAI turned off its Codex models with what Simon Willison called "only giving three days notice". Previews get less still. The 11 retired Gemini API preview models got a median of 28 days, and Gemini 3 Pro Preview got 11 days between its 26 February 2026 deprecation and 9 March shutdown. A developer on Google's forum pointed out that this was less than the two weeks Google's own docs promised previews.

What the providers promise#

Every provider now publishes a lifecycle policy, except xAI. The policies differ in the two numbers that matter: how long a model is guaranteed to stay up, and how far ahead you hear about its end.

Two of these are more predictable than the rest because the end date is known on launch day: Azure's 18 months and Vertex AI's 12-month floor. Mistral's six-month rule for GA models is newer than it looks. The earliest archived copy of its lifecycle page is from 13 August 2026, and every Mistral GA deprecation we counted was announced before then. Only 4 of those 33 retirements gave six months, and the median for 2025 announcements was 29 days.

Dates also move after they're published, usually later. On Vertex AI, Gemini 2.5 Pro's retirement date was 17 June 2026 until at least March 2026, then "not before 16 October", then 16 October, and finally 20 October, the day it was retired. Gemini 1.0 Pro, 2.0 Flash and 2.5 Pro each retired 65 to 125 days after the first date Google posted. OpenAI's gpt-4-0314 had four end dates. Its launch post promised support "until June 14" 2023, and it was finally switched off on 26 March 2026. Extensions help whoever hasn't migrated yet, but they make the dates hard to plan around.

Retirement without an error#

The classic retirement is a hard stop: after the date, calls to the old id fail. A newer pattern keeps the id working and changes what's behind it.

On 8 October 2026 Google deprecated Gemini 3.5 Flash and Gemini 3.7 Flash on the Gemini API, with no notice period. In the changelog's words, "API requests to gemini-3.7-flash are automatically routed to gemini-3.8-flash." Gemini 3.7 Flash had been generally available for 56 days. Code calling it kept running, against a different model. Code that tuned sampling may also behave differently: Gemini 3.6 Flash, where 3.5 Flash requests now go, deprecated the temperature, top_p and top_k parameters.

xAI did the same on 15 May 2026. Eight Grok ids, including grok-3 and grok-4-0709, began redirecting to grok-4.3 after 9 to 17 days' notice. The migration page warns that requests to old ids are billed "at grok-4.3 pricing of $1.25 per 1M input tokens and $2.50 per 1M output tokens rather than the rates of the original model." A forgotten integration changed model and price at once. (We traced how much prices move between model generations in the price of AI tokens.)

For uptime, a redirect is kinder than a 404. But a pinned model id no longer guarantees a fixed model.

One model, several lifespans#

The same model can retire on different dates in different places. ChatGPT retires models much faster than the API, and with far less notice. Cloud resellers keep their own calendars.

ChatGPT's notice is the shortest of any channel. When OpenAI retired GPT-4o from ChatGPT on 13 February 2026, it gave about two weeks; The Register reported OpenAI's figure that only 0.1% of users still chose GPT-4o each day. The API kept serving it. Some providers also soften a retirement after the fact. Anthropic retired Claude 3 Opus from its API on 5 January 2026, then kept it on claude.ai for paid users and available by request on the API.

The weights outlive the endpoint#

A retirement ends one vendor's endpoint, not the model. Where the weights are public, the model keeps running wherever someone hosts it. Anthropic, whose weights aren't public, has committed to preserving the weights of every publicly released Claude model for at least the company's lifetime, but not to serving them.

What this means if you build with AI#

1. Budget a model migration every year or so. Plan for about 18 months on a model, and closer to 13 for one launched recently. A model that's already a year old is in the window where most retirements happen.
2. Pin a dated snapshot, and know its end date. Aliases like gpt-4 or *-latest move under you. A dated id makes the switch your decision. Put the retirement floor (Anthropic's "not sooner than", Azure's 18-month date, Vertex AI's 12 months) in your calendar the day you adopt a model.
3. Watch the deprecation pages, not just your inbox. Notices go by email to active accounts, but the pages are the record. OpenAI exposes a shutdown_date field in its models list and Azure exposes retirement dates in its Models API, so a weekly job can check every model you use.
4. Keep an eval set ready. With a median of 114 days' notice, and 60 at Anthropic, the migration has to fit in a few weeks of engineering time. A saved set of real prompts and expected outputs turns "will the new model break us?" into an afternoon.
5. Log which model actually answered. Redirects (Gemini, xAI) mean a pinned id can start returning a different model's output at a different price. Check the model name in responses, and alert when it changes.
6. Keep previews out of production. Gemini API previews got a median of 28 days' notice, and OpenAI's own page says a preview can go with two weeks'.
7. Weigh notice when you pick a provider. For a workload that's slow to migrate, six months' notice (OpenAI, Bedrock) or a fixed end date set at launch (Azure, Vertex AI) is worth something. So is an open-weight fallback you could host yourself: those weights don't get retired.

The small apps people share on host0 usually call a model their author picked months ago. Every host0 app keeps numbered versions, so when a model retires you can redeploy with the new one and restore the old version in one click if it misbehaves.

Methodology#

We built one dataset of 341 model ids, one row per model per platform, from the providers' own deprecation pages, changelogs, release notes and launch posts, read on 23 October 2026, plus Wayback Machine captures of those pages to date policy changes and moved deadlines. The providers are OpenAI (API, plus ChatGPT and Azure OpenAI in separate tables), Anthropic, Google (Gemini API and Vertex AI), Mistral, xAI, Cohere and Amazon Bedrock. We re-checked 20 headline numbers against the live pages and archive captures before writing.

• Launch is the first day a model id could be called on that provider's API: first public availability for OpenAI (GA or beta), GA for the others. Mistral and xAI launch dates are month-level.
• Retirement is the day the id stopped being served, or started redirecting to another model. Notice is the retirement date minus the announcement date, including retirements that are scheduled but haven't happened.
• The headline set is 204 generally available models with a known launch date on six first-party APIs: OpenAI 105, Claude 29, Gemini API 17, Vertex AI 15, Mistral 33 and xAI 5. It leaves out previews, lab, gated and third-party models, moving aliases, API endpoints, Cohere (no shutdown dates) and Bedrock (a reseller).
• Lifetimes are Kaplan–Meier estimates. Models still live, or deprecated with a future date, count as surviving up to 23 October 2026. The median of retired models alone is 481 days and reads short, because recent models that are still running are left out.
• Eleven OpenAI ids were scheduled to shut down on 23 October 2026, 10 of them in the headline set. We couldn't confirm they did, so we counted them as retired that day; treating them as alive moves the median from 555 to 562 days.
• Limits. OpenAI ids dominate the pool (105 of 204). Our list of OpenAI models still running covers headline models only, which biases OpenAI's lifetime slightly short. Google models appear twice, once per platform. Counting per model id weights big batches; per release batch, the median life is 487 days.

The data pack, with every source, tier, the spot-check log and a do-not-cite list, is in our repository at docs/research/2026-10-23-ai-model-lifespan.md. The merged CSV and the scripts that compute every number in this post are in docs/research/2026-10-23-ai-model-lifespan-data/.

Open questions#

• Did the 23 October shutdowns happen on time? OpenAI's page scheduled 11 model ids for that day, along with the gpt-4 and gpt-3.5-turbo names.
• Why did Anthropic cut notice to 60 days? We found only the documentation edit, no announcement.
• What does a retirement cost? No provider publishes how many customers were still calling a model when it retired, or how long migrations took.
• Will redirects become the norm? Google and xAI now reroute retired ids instead of failing them. Nobody has said how much output changes for the code that never noticed.