Cursor Loses GPT on November 12. Blame the Cap Table.

TL;DR

On August 28, OpenAI said it is winding down the contract that puts its models inside Cursor. Under OpenAI's proposal, Cursor's direct access to OpenAI models ends on November 12, 2026, which is 76 days of notice. No benchmark moved, no price changed, and Cursor did not violate anything. SpaceX closed its $60 billion all-stock acquisition of Cursor in mid-August, and OpenAI says it cannot be confident that Elon Musk's holding company will operate inside its terms of service. If your product's core value is somebody else's model, this is the failure mode you probably have not modeled: not a price hike, not a rate limit, but your own cap table.

What happened, in order

The sequence matters, because each step is boring on its own and only the combination is interesting.

SpaceX first announced the arrangement in April 2026 as an option: roughly $10 billion for a partnership, or $60 billion to buy the company outright later in the year. It exercised the expensive one. The acquisition closed in mid-August, folding Cursor into a group that already contained xAI and X.

Thirteen days later, OpenAI moved.

The clause nobody reads until it fires

OpenAI's stated mechanism is specific and worth understanding, because it is probably sitting in your contracts too. Its custom agreement with Cursor contains a change-of-control provision that gives OpenAI a limited window to cancel after the counterparty is acquired. Windows like that expire. OpenAI used it.

Think of it as the assignment clause in a commercial lease. You did not rent to this person. When the tenant sells the business to somebody you would never have signed with, the clause is what lets you decline to inherit them. The tenant did nothing wrong; the landlord simply never agreed to the new one.

OpenAI also says it held the cancellation to the latest date the contract allowed, which is how you get a 76-day runway instead of a Friday afternoon shutoff. It is a courtesy, and it is also the maximum notice the paperwork permitted. Both things are true.

The reason given is a track record, not a policy

OpenAI did not argue that Cursor is a competitor, which would have been the easy line given that Cursor competes directly with Codex. It argued that it cannot trust the new owner to stay inside the terms of service, and it cited receipts.

• After Musk acquired Twitter, now part of the same group, OpenAI says the company broke the terms of its contract.
• Under oath earlier this year, Musk admitted that xAI distilled OpenAI's models to train Grok, which OpenAI characterizes as a terms-of-service violation.

The second one carries the argument. If your counterparty has testified in court that it already used your models in a way your terms forbid, "we are not confident you will comply" stops being a rhetorical position and starts being a fairly ordinary risk assessment.

Astra is why the bar moved now

The part of OpenAI's reasoning that is easy to skip is the forward-looking one. OpenAI says that as capabilities advance it carries a new level of accountability for how its upcoming model, Astra, gets used. So it is not only cancelling current access; it is declining to ship future models into Cursor at all.

Astra has a history. On August 7, OpenAI disclosed that it could not rule out a Critical cybersecurity capability level for the model, meaning it showed enough agentic coding and security performance that OpenAI could not exclude the possibility it could independently identify and execute attacks against well-defended real systems. That disclosure triggered its Preparedness Framework safeguards, paused some internal work, and pulled in government agencies and outside safety organizations for testing.

So the escalation ladder here is: a model OpenAI is currently too nervous to ship is also the reason it will not hand future models to a company it does not trust. Whatever you think of the reasoning, it is at least internally consistent.

What Cursor still has on the shelf

This is not a product extinction event. Cursor's model picker has been multi-vendor for years, and per Cursor's own docs it currently spans OpenAI, Anthropic, Google, and its own first-party pool.

The quiet detail in that picture: Grok 4.5 now sits in Cursor's own model pool rather than as a third-party option, because the company that makes Grok also owns Cursor. The shelf did not shrink so much as it rearranged around the new landlord.

Cursor also has its own frontier-ish models. Composer has been shipping as the fast default for agentic work for a while, and it is the one product decision that now looks less like cost engineering and more like insurance.

The part that should actually worry you

Forget the personalities. The structural lesson generalizes to anyone shipping on top of a model API, and it is not the lesson most teams have internalized.

The standard supplier risks people plan for are price increases, deprecations, rate limits, and outages. Those are all things you can trigger or absorb. This one is different: the terminating event was a change in Cursor's ownership, and it fired a clause Cursor had presumably signed years earlier. The company got bought for $60 billion and lost a supplier as part of the transaction.

Three concrete things to go check today:

• Read your own change-of-control language. If you take an acquisition, an investment that shifts control, or a majority secondary, which of your model contracts can your suppliers walk away from, and how much notice do you get?
• Know your single-vendor surface. Not "do we support multiple models" but "which flows silently degrade if one provider disappears in 76 days." Evals pinned to one model family are the usual landmine.
• Price the alternative before you need it. Cursor absorbs this because it spent a year building house models. A thin wrapper with no fallback would be a different story entirely.

What is still unknown

Some of this is genuinely unsettled, and it is worth being precise about which parts.

• OpenAI framed November 12 as its proposal under the contract. Whether SpaceX accepts that date or contests the termination has not been reported.
• The announcement addresses Cursor's direct access. Whether users who supply their own OpenAI API key are affected has not been clarified publicly.
• No party has disclosed what share of Cursor usage runs on OpenAI models, so the commercial impact on either side is unquantified.
• SpaceX did not immediately respond to Reuters' request for comment when the story broke.

Key Takeaways

• OpenAI announced on August 28, 2026 that it is ending its model-supply contract with Cursor; under its proposal direct access ends November 12, 2026, giving 76 days of notice.
• The trigger is a change-of-control clause activated by SpaceX's $60 billion all-stock acquisition of Cursor, which closed in mid-August, not by anything Cursor did.
• OpenAI's stated justification is counterparty trust: Twitter allegedly breaking an OpenAI contract post-acquisition, and Musk's sworn admission that xAI distilled OpenAI models to train Grok.
• OpenAI is also refusing to supply future models, citing accountability obligations around Astra, the model it paused on August 7 over a possible Critical cyber capability level.
• Cursor keeps Anthropic, Google, its in-house Composer, and Grok 4.5, which moved into its first-party pool after the acquisition.
• If you build on someone else's models, audit your change-of-control terms now. Your funding or exit event can terminate your supply chain, and no amount of good behavior prevents it.

Sources: OpenAI, "Our decision on Cursor following its acquisition by SpaceX", OpenAI on X, Reuters wire report, August 28, 2026, TechCrunch on the acquisition close, Cursor, "Cursor is now a part of SpaceX", TechCrunch on the Astra pause, Cursor docs, available models, MIT Technology Review on Musk's testimony.