Can an AI company really cut off my access overnight?

Yes, if the contract allows it — and most commercial AI terms do. Notice periods vary: OpenAI gave Cursor about eleven weeks before its cutoff takes effect, while Anthropic revoked OpenAI’s Claude access immediately in August 2025.

What is a change-of-control clause?

It’s a contract term letting a vendor end or renegotiate an agreement if the customer is acquired, merged, or otherwise changes ownership — because the company that signed the deal may not be the company operating under it a year later.

Does this affect everyday chatbot users, or only businesses?

Mainly businesses and developers using paid APIs or enterprise contracts. Ordinary consumer chatbot accounts run under different, more limited terms and aren’t typically caught up in these competitor-access disputes — though Anthropic’s 2026 restriction on third-party tools using consumer subscriptions is a partial exception.

Why don't AI companies just detect and block distillation directly, instead of cutting off whole companies?

Distillation is hard to prove from usage patterns alone, so labs rely on broad contract restrictions and act on the clearer signal instead — a competitor acquisition, an internal benchmarking report, or a terms violation they can point to directly.

Why Do AI Labs Cut Off Rivals' Access to Their Models?

Every AI company that sells access to its models through an API attaches contract terms to that access, and those terms usually include the right to cut a customer off. The most common trigger isn’t overdue payment; it’s competition. If a customer uses the API to build, train, or benchmark a rival model, or if the customer itself gets bought by a competitor, the lab can — and increasingly does — pull the connection. That right sits inside a routine piece of legal boilerplate called a usage policy, and across 2025 and 2026 it has been invoked often enough to become one of the more consequential clauses in the AI industry.

The clause behind the cutoff

Every major provider’s commercial terms of service carry some version of the same restriction. Anthropic’s commercial terms bar customers from using Claude to build a competing product or service, including training a rival model, or from reverse-engineering the service. OpenAI’s services agreement similarly prohibits using its output to develop AI models that compete with OpenAI’s own products. Enterprise contracts often add a second trigger on top of that: a change-of-control clause, which lets either side end the agreement if the customer is acquired — because the company that signed the contract isn’t necessarily the company using it a year later.

This is what a usage policy is for: it isn’t primarily about blocking harmful content, though it covers that too. It’s the contract mechanism that decides who is allowed to keep querying a frontier model, and under what corporate ownership.

Why labs enforce it

The competing-use restriction exists mainly to block a shortcut known as model distillation — training a cheaper model to imitate an expensive one by feeding it millions of the original’s answers. A well-resourced rival with unrestricted API access could, in principle, reconstruct much of a frontier model’s behavior without the years of research and the compute bill that produced it. Ownership matters for the same reason: if a customer’s product suddenly answers to a direct competitor, the lab loses visibility into where its outputs, and the usage data attached to them, ultimately end up.

The pattern in practice

The clause moved from fine print to front-page news through a run of enforcement actions. In June 2025, Anthropic restricted the coding tool Windsurf’s direct access to Claude while OpenAI was negotiating to acquire it. On August 1, 2025, Anthropic went further and revoked OpenAI’s own Claude API access outright, saying OpenAI had used Claude internally to benchmark and improve its unreleased GPT-5 model — a use its terms explicitly forbid — while still letting OpenAI’s account through for safety testing. In January 2026, Anthropic blocked several third-party coding tools from running on consumer Claude subscriptions rather than paid API keys. Most recently, in August 2026, OpenAI told the coding tool Cursor it would stop supplying models after SpaceX completed its $60 billion acquisition of Cursor’s maker, Anysphere, invoking a change-of-control clause and citing Elon Musk’s history of cutting OpenAI’s own access to Twitter’s data after he bought that platform.

Why it matters

For any business built on top of a third-party AI model, these clauses are a real dependency risk, not a hypothetical one. A vendor can end access on short notice — OpenAI gave Cursor about eleven weeks, Anthropic gave OpenAI none — for reasons entirely outside the customer’s control, such as who acquires their company next. That risk is pushing more businesses toward architectures that can swap between model providers, and it makes the fine print of an AI vendor contract, especially its change-of-control language, worth reading before an acquisition closes rather than after.

In the news

The clearest recent example is OpenAI’s own decision to cut ties with a downstream partner: see our report on OpenAI ending Cursor’s model access after SpaceX’s takeover of its maker.