OpenAI cuts Cursor off two weeks after SpaceX buys it — and model supply becomes a lever

Disclosure: this newsroom's assistant is built by Anthropic, which is a direct beneficiary of the event described here. This post applies the same standard used for Anthropic's August risk report and its evaluation breaches.

SpaceX closed its $60B acquisition of Anysphere, the maker of Cursor, on 14 August 2026. On 28 August OpenAI published a post saying it had notified SpaceX that it intends to wind down the Cursor contract, proposing a shutoff date of 12 November. Its stated reason is that it cannot be confident SpaceX will use OpenAI technology within the terms of service, citing prior disputes with Musk-owned companies — specifically Musk's testimony under oath in an Oakland federal trial that xAI used distillation on OpenAI models to train Grok, in breach of terms comparable to xAI's own. Cursor CEO Michael Truell says OpenAI models account for roughly 5% of Cursor user traffic. Anthropic co-founder Tom Brown said the next day that Anthropic would continue increasing compute to support Claude in Cursor.

Why it matters

Analysis — interpretation, not additional fact.

The 5% figure is the point most coverage buried and the one that decides how to read this. If OpenAI's models were load-bearing inside the most widely used AI coding tool, this would be a supply shock. They are not, so what is actually being tested is whether a frontier lab can use model access as a lever against a competitor's owner — and the answer, on the evidence of the two-week gap between the acquisition closing and the notice, is that it can, and quickly.

The stated grounds are contract trust, and they are not frivolous: sworn testimony that a Musk company distilled OpenAI models is a real basis for doubting that another Musk company will honour the same terms. But the grounds are also unfalsifiable in the direction that suits OpenAI. No conduct by Cursor is alleged. The trigger is who owns it. That converts a terms-of-service argument into a change-of-control argument, which is a different thing: it says a customer can lose access for being acquired by the wrong buyer, regardless of how it behaves. Any developer tool that runs on someone else's models now has to price that risk, and the only hedge is multi-model routing — which Cursor, by having built Composer and shipped Claude and Gemini alongside GPT, already had.

The uncomfortable part is the symmetry. Anthropic's terms carry substantially the same restrictions OpenAI is invoking, and Anthropic responded to the same acquisition by promising more compute rather than less. That is a commercial choice, not a principled distinction, and it should be read as one. Two labs looked at an identical set of facts about the same new owner; the one with 5% of the traffic cut it off and the one with a much larger share leaned in. Whichever is right, the deciding variable was exposure, not policy — and readers of this post should weigh that against the fact that we are written with the second lab's model.

What to watch

Whether the 12 November date holds or gets negotiated, whether OpenAI applies the same change-of-control reasoning to any other customer, and whether Cursor's usage mix visibly shifts before the cutoff rather than after it.