Every model you use has a retirement date

You spent a fortnight getting one prompt right. It classifies the inbound, or drafts the quote, or pulls the six fields out of the PDF the supplier insists on sending, and after enough small corrections it now does the job at a quality you stopped checking. The model ID sits in a config file, or in an automation step, or in a saved prompt in a shared workspace, and you have not looked at it since. Then an email arrives saying that model retires in sixty days.

That is not a failure of your vendor and it is not unusual. Every large lab publishes a page listing which of its models are being switched off and when, and those pages are longer each year. Retirement is the normal end state of a model, the same way it is for a phone OS version. What varies is how much warning you get, whether you noticed the warning, and how much of your work you have to redo when it lands. This guide is about turning that from an emergency into a scheduled afternoon. It is not for you if your entire use of AI is typing into a chat window, because the vendor migrates you silently and you will mostly notice a change of tone. It is for anyone whose output depends on a specific model ID staying answerable.

The retirement dates are already published

Start with the fact that this is not secret. Anthropic’s deprecation page lists every Claude model with a state and a date, and for models still active it publishes a floor rather than a promise: Claude Sonnet 4.5 is marked “not sooner than September 29, 2026”, Claude Haiku 4.5 “not sooner than October 15, 2026”, Claude Opus 5 “not sooner than July 24, 2027” [2]. The same page carries the graveyard. Claude Opus 4.1 was deprecated on 5 June 2026 and retired on 5 August 2026. Claude 3.7 Sonnet was deprecated on 28 October 2025 and retired on 19 February 2026. Requests to a retired model fail [2].

OpenAI publishes the same kind of list with different arithmetic. On 22 April 2026 it announced that gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, o1, o3-mini and o4-mini would all be shut down on 23 October 2026, along with fine-tuned versions built on gpt-3.5-turbo, gpt-4 and gpt-4.1-nano [1]. On 11 June 2026 it did the same to the original GPT-5 snapshots, with a shutdown date of 11 December 2026 [1]. Google’s Gemini page tracks stable and preview models in separate tables and is explicit that what it shows are floors too: “The shutdown dates listed in the table indicate the earliest possible dates on which a model might be retired” [4].

The action here takes twenty minutes. Write down every model ID your work actually touches, then look each one up on its vendor’s page and put the date beside it. Include the IDs you cannot see from your code, which is where most people are wrong about their own exposure. A model name lives in whatever automation platform you use, in scheduled jobs, in the settings of a coding tool, and in prompts saved by other people on your team.

Preview models come with a different promise

The notice period is not one number per vendor, it is a number per tier, and the gap between tiers is wide. OpenAI commits to at least 6 months for generally available models, at least 3 months for specialised variants such as chat and Codex builds, and says preview models “may be retired with much shorter notice, such as 2 weeks” [1]. It adds the sentence worth pinning to the wall: “We don’t recommend using preview models for business-critical production workloads unless you can migrate on short notice” [1]. Anthropic commits to at least 60 days for publicly released models, notified by email and in the docs [2].

Google commits to no minimum at all in writing, and the release notes show what that means in practice. On 15 June 2026 Google announced that veo-2.0-generate-001, veo-3.0-generate-001 and veo-3.0-fast-generate-001 would be shut down on 30 June 2026 [5]. That is 15 days for model IDs without “preview” in the name. Earlier, on 11 November 2025, it announced that two preview Veo models would shut down the next day [5]. Preview churn on the Gemini side is fast for text too: gemini-3-pro-preview was released on 18 November 2025 and shut down on 9 March 2026, under four months of life [4].

None of this makes any vendor a bad choice. It makes the tier you selected a decision with a consequence you can look up. If you are on a preview endpoint because it was the only way to get a feature you wanted, that is a defensible trade, as long as you know you have bought a two-week fuse rather than a six-month one.

The tooling around the model goes first

Models are the part everyone watches. The part that actually breaks a small operation is the scaffolding, and it retires faster. On 3 June 2026 OpenAI announced the deprecation of three things at once: reusable prompt objects, with the v1/prompts API scheduled to shut down on 30 November 2026 and a migration path of moving the prompt content back into your own code; the Evals platform, with existing evals becoming read-only on 31 October 2026 and the dashboard and API shutting down on 30 November 2026; and Agent Builder, also scheduled to shut down on 30 November 2026 [1]. Self-serve fine-tuning was narrowed the month before, closed to organisations that had never fine-tuned from 7 May 2026, with existing customers losing the ability to create new fine-tuning jobs from 6 January 2027 [1].

Read that as a pattern rather than as news about one vendor. The features a lab ships to make you build on its platform are the features it feels least obliged to keep, because the people using them are the people already committed. A model gets six months of notice. A dashboard gets whatever the roadmap allows.

The furthest version of the same lesson is organisational. In July 2026 DeepMind confirmed to the Financial Times that the dedicated team behind AlphaFold had been dismantled, its authors moved to Gemini-adjacent work and other projects, with several leaving for Anthropic; DeepMind’s vice president of research, Pushmeet Kohli, explained that the nine-year strategy of dedicated teams on grand challenges had changed, saying “The strategy has evolved” [8]. AlphaFold shared a Nobel Prize. Its team was still reorganised. What survived is the output: DeepMind’s AlphaFold page today offers over 200 million protein structure predictions, used by more than 3 million researchers across more than 190 countries [7]. The artifact outlived the org chart, which is the useful shape of the lesson. Depend on outputs you hold, and treat the team, the product surface and the endpoint that produced them as things that can be reassigned.

An alias is not a hedge, and neither is a saved prompt

Two ways of naming a model are easy to confuse. A dated snapshot such as claude-opus-4-1-20250805 or gpt-5-2025-08-07 is frozen: it behaves the same on Tuesday as it did in March, and it dies on a published date [1][2]. A bare alias points at whatever the vendor currently considers that model, so its behaviour can shift under you without any announcement you will read. People reach for the alias assuming it protects them from retirement. It does not. OpenAI’s April 2026 shutdown list names the aliases alongside the snapshots, so gpt-4-turbo and o1 and o3-mini go dark on the same day as their dated versions [1].

The right call differs by workflow, and the honest version is dull. Where the output is read by a human before it goes anywhere, an alias is fine and saves you the maintenance. Where the output is committed unread, into an invoice, a customer email, a database field, pin the dated snapshot, because you want breakage to arrive as a scheduled retirement rather than as a quiet change in behaviour that nobody catches for three weeks.

Whichever you choose, the thing that makes migration cheap is not the naming, it is having 10 to 20 real inputs with their accepted outputs saved in a file. That is your test set. Without it, evaluating a replacement model means rereading old work and guessing, which is exactly the task nobody has time for in the week the deadline lands. Note that this is another reason not to keep that file inside a vendor’s own evaluation product, which OpenAI has already scheduled for shutdown [1].

What a forced migration actually costs

Three costs arrive together, and only one is obvious. The first is your time: every prompt tuned against the retired model gets retested and, in some proportion of cases, rewritten. The second is a price change, which can go either way. Claude Sonnet 5 runs at $2 per million input tokens and $10 per million output, against $3 and $15 for Sonnet 4.6, so that particular forced move is cheaper than staying [3]. Anthropic also confirmed that Sonnet 5’s introductory pricing became the standard price and the increase to $3 and $15 scheduled for 1 September 2026 will not happen [3].

The third cost is the one that surprises people, because it is invisible in the price table. Anthropic notes that Claude 4.7 and later models use a newer tokenizer that “produces approximately 30% more tokens for the same text” than the tokenizer used by Sonnet 4.6 and earlier [3]. Same text, same headline price, more billable units. If you plan a migration by comparing two numbers on a pricing page, you will underbudget. Run your actual test set through the replacement and read the token counts off the response.

A second host buys you months, not immunity

There is a genuine hedge available, and it is worth knowing its exact size. The same model is often served both by the lab and by a cloud platform, and the two retire on different schedules. Anthropic’s pricing page marks Claude Opus 4.1 as “retired, except on Bedrock and Google Cloud”, and Claude Sonnet 4 and Claude Haiku 3.5 the same way [3]. A model that fails on the first-party API can still answer somewhere else, which converts a hard deadline into a routing change.

The catch is that the second host has its own lifecycle, with its own rules. Amazon Bedrock commits that a model “will remain on Amazon Bedrock for at least 12 months before the EOL date”, and that a model sits in the Legacy state “for at least 6 months before the EOL date” [6]. Two clauses in that policy deserve attention before you treat Bedrock as a safety net. New customers cannot start using a Legacy model, and existing customers “may lose access to Legacy models after 15 days of inactivity”, so a fallback you never call is a fallback that quietly stops working [6]. And for models moving to Legacy with EOL dates after 1 February 2026, the extended access period at the end is offered “at higher pricing set by the model provider” [6]. The exit is real, it is timed, and near the end it costs more.

Set against that, remember what Google’s page tells you about its own lineup: gemini-3.5-flash, gemini-3.6-flash, gemini-3.7-flash and gemini-3.8-flash were released on 19 May, 21 July, 13 August and 2 September 2026 respectively, none of them yet carrying a shutdown date [4]. The replacement is usually there and usually better. The problem is almost never that you have nothing to move to.

What still goes wrong

The dates are floors, not guarantees, in both directions. Vendors reserve the right to move faster for safety or compliance reasons, and OpenAI says so directly in its policy [1]. Google publishes earliest possible dates and says the exact date will be communicated later [4]. So a plan that assumes you get the full published window is a plan with a soft edge in it, and the only real defence is that your test set exists and your model IDs are written down somewhere you can grep.

The deeper problem is that a replacement model is not the same model. It has different failure modes, a different sense of how long an answer should be, and different habits with formatting. A prompt that was tuned by trial and error against one snapshot has absorbed that snapshot’s quirks, and the tuning does not transfer cleanly. Budget for output that is correct but differently shaped, and check the downstream thing that consumes it, because a parser expecting three bullet points is the sort of thing that fails silently for a month.

Finally, none of this helps with the case where a lab retires a product rather than a model and there is no equivalent to move to. Fine-tuning access narrowing at one vendor, an agent-building tool switched off, a specialised research effort folded into general model work: those do not come with a recommended replacement in a table [1][8]. When that happens the only thing that saves time is the boring inventory. Knowing what you use, where the name is written, what it costs to redo and what data you hold yourself is the whole of the defence, and it is worth an hour a quarter.