What actually binds an AI vendor

You have a few Zapier steps, a Notion database, a script on a cron job and a folder of saved prompts, and every one of them calls a model by name. It works, so you stop looking at it. Then an email arrives saying that model retires in 60 days, and you find out which parts of your business were load-bearing.

Most advice about choosing an AI supplier points at governance. Who controls the board, whether the nonprofit still has teeth, what the trust is for. That material is real and it takes 10 minutes to read, but it is not a promise to you, and there is nothing in it you can put in a calendar. The documents that actually constrain a vendor’s behavior toward your account are duller and more specific: the deprecation page, the data-use policy for your tier, and the version names you are calling. This guide is for one-to-ten-person operations buying off the public price list. If you have a procurement team and a negotiated master agreement, your leverage is different and your lawyer is the better guide.

Governance structures tell you about intent, not about your account

Anthropic’s Long-Term Benefit Trust has five Trustees. At the close of the Series C, Anthropic amended its corporate charter to create a new class of stock, Class T, held exclusively by the Trust. That stock carries authority to elect and remove a number of Anthropic’s board members, phasing in according to time-based and funding-based milestones, and the company states the Trust will elect a majority of the board within 4 years. Anthropic says the Trust’s arrangements are designed to insulate the Trustees from financial interest in the company. The structure was announced on 19 September 2023 [4].

OpenAI took a different route. On 28 October 2025 it recapitalized. The nonprofit, now called the OpenAI Foundation, controls the for-profit business, holds equity in it currently valued at approximately $130 billion, and gains additional ownership as the for-profit reaches a valuation milestone. The commercial entity is a public benefit corporation called OpenAI Group PBC. The Foundation said it will initially focus on a $25 billion commitment across two areas, health and curing diseases, and technical solutions to AI resilience [5].

Both structures are genuine and both are unusual. Neither is a commitment to you. Read the two pages and you will not learn when your model goes away, whether your prompts are used for training, or what happens to your account if the company’s priorities shift. A governance structure is a statement about who decides. The documents below are statements about what the vendor owes the person paying.

The data clause you are under depends on the tier you pay for

The same brand gives different answers depending on which door you came in through, and the marketing page is not the door.

Anthropic states that by default it will not use inputs or outputs from its commercial products, naming Claude for Work, the Anthropic API and Claude Gov, to train its models. The exceptions are narrow and opt-in: feedback you submit through the thumbs up and thumbs down buttons, bugs you report, or cases where you otherwise choose to allow it. When you submit feedback, the related conversation is stored for up to 5 years. Team and Enterprise plan owners can switch off feedback entirely using the “Rate chats” setting under Organization settings. Consumer plans are governed by a separate policy [7].

OpenAI states that by default it does not use business data to train its models, and that it retains API inputs and outputs for up to 30 days to provide the services and identify abuse, after which the data is removed unless it is legally required to keep it. Zero data retention can be requested for qualifying use cases. It lists a completed SOC 2 Type 2 audit across ChatGPT Enterprise, Edu, Healthcare, Business and its API platform [6].

Google’s Gemini API terms draw the line at payment. On the unpaid tier, Google uses the content you submit and the generated responses to provide, improve and develop Google products and services, human reviewers may read, annotate and process your inputs and outputs, and the data is disconnected from your Google Account, API key and Cloud project before reviewers see it. On the paid tier, Google does not use your prompts or responses to improve its products. Separately, for Grounding with Google Search, Google stores prompts, contextual information and output for 30 days [8].

The practical version: find the policy page that names your exact plan, and read that one. A free key you added to a Make scenario 8 months ago is on different terms from the paid seat you use in the browser [8].

Model retirement is the vendor risk that actually reaches you

Anthropic has committed to preserving the weights of all publicly released models, and all models deployed for significant internal use going forward, for at minimum the lifetime of Anthropic as a company. It has also committed to producing a post-deployment report when a model is deprecated, and to interviewing the model about its own development, use and deployment, recording the responses, with particular care to document any preferences it has about the development and deployment of future models [3].

That is a commitment about preservation. It is not a commitment about availability. A retired model is one you can no longer call, whether or not its weights sit on a disk somewhere. Anthropic’s own table shows claude-opus-4-1-20250805 deprecated on 5 June 2026 and retired on 5 August 2026, and claude-3-7-sonnet-20250219 deprecated on 28 October 2025 and retired on 19 February 2026 [1].

OpenAI’s page shows the same mechanism at a larger scale. Legacy snapshots including gpt-3.5-turbo and gpt-4 were announced on 22 April 2026 for shutdown on 23 October 2026. Older GPT-5 and o3 snapshots were announced on 11 June 2026 for shutdown on 11 December 2026, with gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna named as replacements. Transcription models including whisper-1 and gpt-4o-transcribe were announced on 26 August 2026 for shutdown on 26 February 2027 [2].

None of this is a vendor behaving badly. It is the normal metabolism of the product. The failure mode is not the retirement, it is finding out about it from an outage.

Pin the version, then put the retirement date in your calendar

Both vendors hand you a calendar for free, and almost nobody uses it.

Anthropic publishes a tentative retirement date for models that are still active, before any deprecation is announced. As of today the table lists claude-sonnet-4-5-20250929 as not retiring 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, and claude-fable-5-1 not sooner than 1 September 2027 [1]. Those are floors, not schedules, and a floor 24 days away is still useful information about where you are standing.

To use them you need to know which version you are actually on, which means calling dated snapshot names rather than floating aliases in anything that matters. An alias quietly moves you to a new model and your outputs change without a deploy. A snapshot name breaks loudly on a date you were told about in advance. Loud and scheduled beats quiet and whenever.

Then write down every place a model name appears. In a small operation that list is longer than you expect, because model names hide in dropdowns: a step in Zapier, a node in n8n, a module in Make, a connected assistant in Notion, plus the two scripts and the browser extension. The list is the deliverable. Once it exists, a retirement notice becomes an hour of work instead of an afternoon of archaeology.

Rehearse the switch before the notice arrives

A notice period is only worth what you can do inside it, and the thing that takes the time is not swapping a string. It is discovering that you never wrote down what “working” looked like.

Keep the material outside the vendor: the prompt text, 10 to 20 real inputs, and the outputs you were happy with. A plain folder or a Notion database is fine. That file is the only thing that lets you tell, in an afternoon, whether Claude, ChatGPT or Gemini does the job you had built. Without it you are re-deriving your own requirements under time pressure, which is how people end up staying on a worse option because moving felt risky.

Run the switch once while nothing is wrong. Take the 10 tasks that matter, run them against a second vendor, and write down where the output degraded. You are not looking for a winner. You are looking for the specific sentence in your prompt that only worked because of one model’s habits, because that sentence is where the next migration will hurt.

The notice period is the number to compare vendors on

If you check one document before committing a process to a vendor, check this one, because it is the only place the vendor writes down a number it owes you.

OpenAI publishes minimum notice periods before model retirement: at least 6 months for generally available models, at least 3 months for specialized variants such as chat, Codex and deep research variants, and much shorter notice for preview models, such as 2 weeks. It adds that if safety or compliance concerns require retiring a model sooner, it will provide as much notice as reasonably possible [2]. Anthropic notifies customers with active deployments of upcoming retirements, providing at least 60 days’ notice before model retirement for publicly released models [1].

The gap is large and it shows up in practice, not just in the policy text. Anthropic’s published dates cluster on its own floor: claude-3-haiku-20240307 was deprecated on 19 February 2026 and retired on 20 April 2026, exactly 60 days; claude-opus-4-20250514 ran from 14 April 2026 to 15 June 2026; claude-opus-4-1-20250805 from 5 June 2026 to 5 August 2026 [1]. OpenAI’s December 2026 shutdown was announced on 11 June 2026, exactly 6 months ahead [2].

Note which way that cuts. The vendor whose board is elected by a trust built to insulate its trustees from financial interest in the company [4] gives you 60 days. The vendor that recapitalized around a nonprofit holding roughly $130 billion of equity [5] gives you 6 months. Governance posture and customer notice are separate variables, and only one of them changes what lands in your inbox.

What still goes wrong

Every number here is policy, not contract. OpenAI’s notice periods carry an explicit exception where safety or compliance concerns require retiring a model sooner [2], and a published policy page can be edited without anyone telling you. The floor is a floor on ordinary behavior, not a guarantee. Treat these figures as the vendor’s stated intent expressed precisely enough to plan around, which is more than a governance page gives you and less than a signed term.

The comparison also flatters whichever vendor writes the clearest page. Anthropic’s 60 days is worse than OpenAI’s 6 months, but both publish a number and a dated table, which puts them ahead of any tool whose model provider you cannot even identify. If you route through a wrapper, an agency build or a plugin, the notice period that matters is the one your intermediary passes on, and that is usually zero because nobody thought to write it down.

Finally, none of this tells you which vendor is better, and it is not meant to. A model can be retired on 6 months’ notice and still be the right thing to build on, and a company with immaculate governance can ship a version that quietly breaks the one prompt you depend on. This guide is not for anyone whose obligations require negotiated terms, audit rights or a contractual notice period; if that is you, the public pages are the starting point for a redline, not the answer.