How to depend on an AI lab you don't control

You chose a model once. Someone compared three of them over an afternoon, one won, and the name of that winner is now hard-coded in a script, typed into an automation step, and written into onboarding notes nobody has reopened since. In the months after that afternoon the vendor renamed its lineup, shipped two models above the one you picked, published a date on which yours stops answering, cut the price of the newer model below the price of yours, and replaced some of the people who built it. None of that arrived as an email you acted on.

This guide is for solo operators and small teams who send work to a hosted model API or a paid assistant and would rather not be surprised by any of that. It is not for anyone with an enterprise agreement and a named account rep, whose notice periods and pricing are negotiated rather than published. It is not for teams running open weights on their own hardware, who have traded this problem for a different one. Every date and price below was read from the vendor’s own page on 5 September 2026, and all of them can move.

The people who build your model can leave in a single morning

On 5 August 2026 it was reported that four senior Google researchers were leaving the company together to start their own venture. Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals left to found Discovery Loop, a public benefit corporation aiming to use AI to accelerate scientific research by automating experimental processes and exploring recursive self-improvement, with Dean as chief executive [8]. Ghemawat was a senior fellow, Le was a founding member of Google Brain, and Vinyals had been a senior research scientist at Google DeepMind [8].

The useful reading of that is not a forecast about Google, and anyone selling you one is guessing. The useful reading is structural. The delivery date you were quoted for the next model, in a keynote or a podcast or a sales call, is produced by people who are free to resign, get reorganised, or ship something else instead. You have no contract with any of them. What you do have is the vendor’s own documentation, and it is worth noticing what that documentation is willing to promise. None of the three deprecation pages names a date on which a future model arrives. All three name, in a table, the date the current one stops answering [1][2][3].

That asymmetry is the whole planning problem. Arrival is marketing and removal is policy. Build on the second one.

The retirement date is the only date the vendor puts in writing

Anthropic runs a four-state lifecycle. Active means the model is fully supported and recommended. Legacy means it will no longer receive updates and may be deprecated later. Deprecated means it still works but is no longer recommended, with a replacement named and a retirement date assigned. Retired means requests to it fail [1]. The company says it notifies customers with active deployments and gives at least 60 days’ notice before retiring a publicly released model [1]. Its table then puts a tentative floor under each model: claude-sonnet-4-5-20250929 not sooner than 29 September 2026, claude-haiku-4-5-20251001 not sooner than 15 October 2026, claude-opus-4-5-20251101 not sooner than 24 November 2026, claude-opus-5 not sooner than 24 July 2027 [1]. The floors are real. claude-opus-4-1-20250805 was deprecated on 5 June 2026 and retired on 5 August 2026, with claude-opus-4-8 named as the replacement [1].

OpenAI’s commitment is longer and graded by how finished the model is. Generally available models get at least 6 months of notice before shutdown, specialised variants at least 3 months, and preview models, identified by the word preview in the model name, may be retired with much shorter notice, such as 2 weeks [2]. Safety or compliance concerns can compress any of that, with the company saying it will give as much notice as reasonably possible [2]. The dated entries look like the Anthropic ones: the gpt-5-2025-08-07 snapshot shuts down on 11 December 2026, with gpt-5.6-sol named as the replacement [2].

Google is the loosest of the three, and says so plainly. Its deprecations page states that the shutdown dates listed in the table indicate the earliest possible dates on which a model might be retired, and that it will communicate the exact shutdown date to users with advance notice [3]. Read that twice before you plan around it. A date you cannot rely on being early or late is not a date, it is a warning. The same page carries gemini-3.1-flash-lite with a 7 May 2027 shutdown and gemini-3.5-flash-lite as the replacement [3], and leaves most of the current lineup with no shutdown date announced at all [3].

Three practical consequences. Anything labelled preview can be pulled on as little as 2 weeks’ notice under OpenAI’s published terms [2], which makes it fine for a prototype and wrong for anything a client sees. The day you pin a model version, put its floor date in the same calendar you use for tax deadlines, because the notice goes to customers with active deployments [1] and not necessarily to the person who wrote the script. And check the page rather than waiting for the notice, since all three vendors maintain these tables in public [1][2][3] and reading one does not depend on a message reaching you.

The price you built on is a rate with an expiry date

Prices move in both directions and the vendors publish the moves in advance, which means you can read them and almost nobody does. Gemini 3.8 Flash is listed at $0.75 per 1 million input tokens and $3.75 per 1 million output tokens through 31 December 2026, then $1.50 and $7.50 starting 1 January 2027 [4]. That is a doubling, dated, sitting on the public price page today. A workload with thin margins at the first number is underwater at the second, and the change will not announce itself in your invoice until January.

The other direction is just as common and costs you quietly. On Anthropic’s list, Claude Sonnet 5 is $2 per million input and $10 per million output, while the older Claude Sonnet 4.6 is $3 and $15 [5]. Anything still pinned to the older identifier is paying 50 percent more than the newer Sonnet on the same list, for a model whose retirement floor arrives more than four months earlier, on 17 February 2027 against 30 June 2027 [1]. Pinning is the correct call for reproducibility; forgetting the pin for a year is not. The gaps across vendors are wider than the gaps within one. At the low end, Gemini 3.5 Flash-Lite runs $0.30 and $2.50 [4] against Claude Haiku 4.5 at $1 and $5 [5], so a call carrying 6,000 input and 1,000 output tokens costs about $0.0043 on the first list and about $0.011 on the second, roughly 2.5 times more.

Two discounts are already on the price lists and neither requires a conversation. Anthropic’s Batch API applies a 50 percent discount on both input and output tokens for asynchronous work, which puts Sonnet 5 at $1 and $5 [5], and a cache read bills at 0.1 times the base input price [5]. Batch pricing matters here for a reason beyond the invoice: it makes the periodic re-test described below cost roughly half of what it otherwise would.

Portability is a shape you keep, not a tool you buy

The mechanical part of switching providers is now close to trivial, and the vendors did that work themselves. Anthropic documents using the OpenAI SDK against the Claude API by changing three things: the API key, the base URL to https://api.anthropic.com/v1/, and the model name to something like claude-opus-5 [6]. Google documents the same pattern against https://generativelanguage.googleapis.com/v1beta/openai/ with a Gemini API key and a model such as gemini-3.8-flash [7].

The interesting part is what both pages then admit. Anthropic says the compatibility layer is primarily intended to test and compare model capabilities, and is not considered a long-term or production-ready solution for most use cases [6]. On that layer the strict parameter is ignored, so tool-use JSON is not guaranteed to match your schema; audio input is ignored; prompt caching is unavailable; multiple system messages are hoisted and concatenated into one; and logprobs, response_format, seed and logit_bias are among the fields silently ignored rather than raising an error [6]. Google’s page says support for the OpenAI libraries is still in beta while feature support is extended, and that parameters outside its documented set are silently ignored by the compatibility layer [7]. Silently is the operative word in both. Your code will not break. Your output will change.

So the honest version is this. A plain text-in, text-out call is a config change. Anything using thinking, prompt caching, structured output guarantees, citations or document handling is a rewrite, because those are the features each vendor built to be sticky. The cheap insurance is architectural rather than commercial: put every model call behind one function in one file, keep the prompt text in files rather than in string literals, and keep any vendor-specific flag in the same place as the call. Then a switch is a diff in one file rather than a search across a codebase. Anthropic also notes that a prompt heavily tweaked for one vendor is likely well-tuned to that vendor specifically and worth reworking for the other [6], which is the part no abstraction layer fixes.

A second provider you have never run is a bookmark, not a backup

The failure that actually hurts is not the deprecation notice. It is getting the notice with 60 days on the clock [1] and discovering that nobody knows whether the replacement model does the job, because the only evidence you ever had was the afternoon someone spent comparing three of them a year ago.

Fix it with the smallest possible test set. Take 20 to 30 real inputs from the workload, the messy ones included, and write down the output you would actually accept for each. That is your evaluation set, and it answers a question no published benchmark is asked, which is whether the model does your work. Run it against the current model, the vendor’s named replacement, and one model from a second provider. Compare the outputs yourself, price all three with the calculator above, and write the result into a file with the date on it. Then repeat when a deprecation appears on any of the three pages [1][2][3], and once a quarter regardless.

The point of the exercise is not to switch. Most of the time you will keep what you have. The point is that on the day you need to move, moving becomes a decision you already have data for, taken in an afternoon, rather than a research project run under a deadline the vendor set.

What still goes wrong

Portability caps out exactly where the value is. The features that make a vendor worth using, prompt caching, extended thinking, citations, document handling, guaranteed structured output, are the features the compatibility layers drop or ignore [6][7], and Anthropic describes its own layer as a way to test and compare rather than a production solution for most use cases [6]. Genuine two-provider redundancy therefore means writing to two native SDKs and maintaining two sets of prompt text, which is real work with a real ongoing cost. For most small teams the correct answer is not redundancy but reversibility: one provider, one round of structural tidiness, one tested exit.

The notice periods are minimums and they have carve-outs. OpenAI’s commitment is at least 6 months for generally available models but explicitly allows shorter timelines where safety or compliance requires it [2], and Google describes its shutdown dates as the earliest possible dates rather than the actual ones [3]. A minimum is a floor, not a plan, and a date that can only move in one unspecified direction is not something to schedule a launch around.

None of this touches the subscription side. If your team works inside a chat assistant rather than an API, you get the model the vendor gives you on the day, with no pinning, no version table, and no notice beyond a changelog. That is a reasonable trade for most work. It is worth knowing you have made it, and worth keeping the one workflow that genuinely cannot change output shape on an interface where the version is something you choose.