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

You built something on one AI vendor. Client research, first drafts, support replies, a few automations that run without you watching. The failure you planned for is the obvious one: the model gets worse, or the bill goes up, and you grumble and adjust. The failure most operators have not planned for is losing access for a reason that has nothing to do with anything you did.

That second kind is what this guide is about, and the fix is not paranoia. It is thirty minutes of reading and one rehearsal a quarter. This is written for a solo operator or a small team with one API key, a credit card and no procurement department. It is not legal advice, and it is not for anyone running a negotiated enterprise deployment with counsel and a security review attached; those people have contract levers you do not have. What you have instead is the ability to leave quickly, which is worth more than it sounds if you keep it in working order.

The document that ends your access is not the one you signed

Two documents govern your relationship with an AI vendor, and most people have read neither. The terms of service set out the commercial relationship. The usage policy, sometimes called the acceptable use policy, sets out what you are allowed to do, and it is the one that gets you cut off.

The terms are blunter than the marketing. Anthropic’s Consumer Terms of Service, effective 8 October 2025, state: “We may suspend or terminate your access to the Services (including any Subscriptions) at any time without notice to you if we believe that you have breached these Terms, or if we must do so in order to comply with law” [2]. The same document covers what happens to your history: “Upon termination of these Terms, a Subscription, or your access to the Services, we may at our option delete any Materials or other data associated with your Account” [2]. OpenAI’s usage policies, effective 29 October 2025, say the same thing more briefly: “breaking or circumventing our rules and safeguards may mean you lose access to our systems or experience other penalties” [5]. Neither vendor is being unusual here. This is standard, and it is standard because it has to be.

The usage policy is also a moving target. Anthropic’s current one is effective 15 September 2025 and carries a “Previous Version” link at the top, which is a polite way of saying the text you are bound by is not the text you skimmed when you signed up [1]. It contains obligations that catch ordinary small businesses, not just bad actors. For what it calls High-Risk Use Cases, covering legal, healthcare, insurance, finance, employment, housing, academic testing and journalism contexts, the policy requires that “a qualified professional in that field must review the content or decision prior to dissemination or finalization”, and that you disclose AI involvement to the people affected “at a minimum at the beginning of each session” [1]. If you draft employment screening notes, tenant summaries or insurance explainers for clients, that is a rule you are already subject to.

Your vendor’s policy can be renegotiated by people who are not you

The policy you read is a default, and defaults get bargained over above your head. Anthropic’s usage policy says so directly: “Anthropic may enter into contracts with certain governmental customers that tailor use restrictions to that customer’s public mission and legal authorities if, in Anthropic’s judgment, the contractual use restrictions and applicable safeguards are adequate to mitigate the potential harms addressed by this Usage Policy” [1]. OpenAI’s version of the same idea is a gate rather than a carve-out: it prohibits use for “national security or intelligence purposes without our review and approval” [5]. Either way, a large enough customer gets a different rulebook than the one on the public page.

The pressure runs the other way too, and it can take your access with it. On 30 July 2026, US District Judge Rita Lin ruled that the government still lacked the evidence to justify designating Anthropic a supply-chain risk, the label behind a ban on federal use of its models. She had already blocked that ban temporarily in March and was deciding whether to make the block permanent [8]. The government had argued that Anthropic’s public criticism of the DOD justified the ban, reasoning Lin called “really troubling” because of the precedent it would set for retaliating against contractors who disagree with the administration. The DOD also claimed Anthropic could disable or alter models during operations; Lin said she saw no proof the company could alter a delivered model or “flip some kind of kill switch” [8].

Set aside who is right. The durable fact for you is the shape of it: a very large buyer attempted to make one vendor unavailable to an entire class of users, over a disagreement that none of those users were party to. That specific case will resolve one way or the other, and it will stop mattering. The pattern will not stop. Copyright litigation, export controls, state AI statutes, a payment processor’s risk team, an app store review, a jurisdiction added to or removed from a supported-countries list. Your access sits downstream of arguments you cannot join.

Model retirement is the version of this you can put in a calendar

Strip out the politics and the same dependency shows up on a published schedule, which makes it the cheapest possible rehearsal. The model you use today will be switched off, and the vendors tell you roughly when.

Anthropic commits to “at least 60 days’ notice before model retirement for publicly released models” [3]. In practice that floor is the plan, not a cushion: claude-opus-4-1-20250805 was deprecated on 5 June 2026 and retired on 5 August 2026, a gap of 61 days [3]. The current table lists claude-sonnet-4-5-20250929 as retiring not sooner than 29 September 2026 and claude-haiku-4-5-20251001 not sooner than 15 October 2026 [3]. If either of those strings is sitting in one of your automations right now, you have weeks, not quarters.

OpenAI publishes longer floors and more of them: at least 6 months for generally available models, at least 3 months for specialized variants such as chat and Codex variants, and “much shorter notice, such as 2 weeks” for preview models [4]. It also tells you plainly not to build on the short-notice tier, advising against preview models for business-critical production workloads unless you can migrate quickly [4]. Recent entries give the texture: transcription models including whisper-1 were announced as deprecated on 26 August 2026 with removal from the API on 26 February 2027 [4].

The instruction here is small and boring. Find every place a model name is written down, in a script, an n8n or Zapier step, a saved prompt, and look up its retirement date today. Then treat the next forced migration as a free fire drill rather than an emergency, because it is the one version of this problem that arrives with a date attached.

What switching actually costs

The wiring is the easy part, and knowing that changes how you think about the rest. Anthropic publishes a compatibility layer that lets the OpenAI SDK talk to Claude by changing three things: the API key, the model name, and one line setting base_url to https://api.anthropic.com/v1/ [6]. The vendors converged on a similar request shape, which is why a swap that ought to be a rewrite is three lines.

Anthropic is also honest that it is not a destination: “This 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]. Running through it, prompt caching is unavailable, the strict parameter for function calling is ignored, and system messages get concatenated together [6]. Good enough to prove a fallback works in an hour. Not good enough to live in.

So the real cost sits everywhere else. Prompts tuned over months to one model’s habits, which stop landing the same way. Tool and function schemas that need rewriting. Whatever you use to tell good output from bad, which for most solo operators is a feeling rather than a file, and feelings do not transfer. Then the tail: a week or two after the switch where the work is measurably worse and you are correcting more of it. That tail is usually the largest number in the whole exercise and the one nobody writes down.

Keeping a second vendor warm

Name a specific fallback, in writing, before you need one. Not “we could always move to ChatGPT”. A vendor and a model name, chosen because you have actually run your work through it.

Then run that work. Take your three highest-value jobs, the ones that would hurt most to lose, and push them through the second model. Keep both sets of outputs in a file so you can compare rather than remember. This costs very little: Gemini’s API has a free tier listed as “Free of charge”, and on the paid tier Gemini 3.5 Flash-Lite runs $0.30 per million input tokens and $2.50 per million output [7]. A serious portability test for a few dollars is not a budget conversation.

Get your working material out of the chat window while you are at it. Prompts, reference documents and the handful of examples you use to sanity-check quality belong in files you control, a repo or a Notion database or a plain folder, not in a conversation history you cannot export cleanly and that the terms permit the vendor to delete at its option on termination [2]. A prompt library that lives in one vendor’s UI is a dependency wearing a disguise.

Last, put it on a schedule. Thirty minutes a quarter, or piggyback it on the next model retirement date you looked up, since that migration is happening anyway.

What still goes wrong

A second vendor is not independence when both vendors forbid the same things. Anthropic and OpenAI publish heavily overlapping prohibited-use categories, covering weapons, cyber intrusion, fraud, surveillance and automated high-stakes decisions without human review, so a use that trips one will usually trip the other, and no amount of portability saves you from a rule you are genuinely breaking [1][5]. Portability protects you against a commercial decision, a political fight or a retired model. It does not protect you against your own product being outside the lines. That means the usage policy is worth reading as a product specification, once, properly, before you sell the thing.

The notice periods also have escape hatches, and the escape hatches are where the bad days live. OpenAI’s floors apply “Unless safety or compliance concerns require a faster timeline” [4], and Anthropic’s suspension clause is explicitly “without notice” [2]. The version of this problem that arrives with 60 days and a migration guide is the polite version. The other version arrives on a Tuesday morning with a locked account and a support queue, and the only thing that helps then is work you did months earlier.

Finally, none of this resolves the underlying uncertainty, and you should not build as though it will. As of the ruling described above, the permanent block was still undecided, and that case was one of two suits Anthropic filed in March 2026, with the second heard separately in Washington [8]. By the time you read this, that one may have landed. It will not change the calculation, because nobody gets to know how the next one ends before they have to make decisions. The point of a rehearsed fallback is that you stop needing to know.