OpenAI will terminate its contract with Cursor on November 12, 2026, invoking a clause that lets it walk away within a limited window after a change of ownership. The trigger was SpaceX's acquisition of Cursor. OpenAI says it cannot be confident SpaceX will comply with its usage terms, and that November 12 is the longest notice the agreement permits. Cursor co-founder Michael Truell says GPT models accounted for only about 5% of the company's AI-related traffic. Anthropic moved to pick up the slack.
Cursor users are not fully cut off. Anyone with their own OpenAI API key can keep calling GPT models through the editor, and OpenAI will continue to serve its models through its own extensions inside development environments. What ends is the commercial relationship between the two companies.
OpenAI spokesperson Thibault Sottiaux framed the decision as a question of trust, and was careful to say it is aimed at Elon Musk's companies rather than at coding tools as a category or at the developer community around them. That distinction matters to OpenAI's positioning and not much to Cursor's customers, who experience the same outcome either way.
The grievance behind it is older than the acquisition. OpenAI accuses Musk of breaking contracts after he bought Twitter and of trying to obtain data outside the terms of their agreements. The lab held a licensing deal with Twitter giving it access to the full tweet firehose for roughly $2 million a year, used to train ChatGPT. In December 2022 Musk learned of the arrangement, decided the price was too low, and cut OpenAI's access to the data. During litigation Musk also acknowledged using data from other AI labs' models to train his own Grok models — distillation, which likewise violates OpenAI's terms.
OpenAI ties the move to safety as well as contracts. Its forthcoming Astra model is expected to have advanced cybersecurity capabilities, which the company says raises the bar for how rigorously partners must observe usage terms. That reasoning has an awkward shape. OpenAI has spent years warning about AI risk; saying that a new model demands a new level of partner responsibility implies the previous level of control was adequate for everything shipped so far, which is not a claim the company would make in any other context.
Truell responded quickly. He put OpenAI's share of Cursor's AI traffic at roughly 5%, said the companies are in talks to find a resolution, and pointed out that Cursor was one of OpenAI's earliest customers. The two teams worked closely for years, he said, and Cursor built its business on the assumption that OpenAI's platform was neutral infrastructure.
Anthropic used the opening. Tom Brown, co-founder and chief compute officer, wrote on X that Cursor has been a reliable partner since Sonnet 3.5 and that Anthropic will keep adding compute for Claude models inside Cursor. It is a well-timed message with a short memory attached. Anthropic cut Windsurf off from Claude in June 2025 after reports that OpenAI intended to buy it, and in August 2025 revoked OpenAI's own API access over reports that OpenAI teams had used Claude for internal benchmarking ahead of the GPT-5 launch. Both cutoffs were triggered by the same thing triggering this one: who owns the customer.
The number that deserves the most attention is Truell's 5%. If accurate, it says the model layer beneath a coding product has become substitutable to the point where losing an entire frontier vendor is a routing change. That is a worse outcome for OpenAI than for Cursor, and it undercuts the theory that model access is the leverage point in this market. Read alongside the fact that OpenAI still ships its own IDE extensions, the termination looks less like a safety measure and more like the removal of a distribution partner that is now owned by a competitor and was, on its own numbers, barely buying anything.
What is notably absent from OpenAI's reasoning is any allegation that Cursor or SpaceX has violated anything. The clause being exercised is triggered by ownership, not conduct. The evidence of bad faith on offer is Musk's behavior at Twitter in 2022 and his admissions about Grok's training data — neither of which involves Cursor, a company OpenAI supplied without complaint until SpaceX bought it. Model access is being withdrawn on the basis of who signs the acquirer's paychecks, which is a defensible commercial position and a different thing from the trust-and-safety argument being made for it.
Truell's phrase — neutral infrastructure — is the part worth keeping. Every application company building on a frontier lab is building on a supplier that also sells a competing product and reserves the right to leave when the cap table changes. The change-of-control clause is no longer boilerplate; it is now a term that prices acquisitions, and any startup taking an offer from the wrong buyer has to assume its model access is part of what gets sold.