SAN FRANCISCO, 30 AUG 2026 — OpenAI will end direct access to its models inside the Cursor code editor on 12 November. The move follows SpaceX's acquisition of Cursor's parent company Anysphere, with OpenAI invoking a change-of-control clause in its contract. Cursor's chief executive says OpenAI models handle about 5 per cent of its traffic.

What OpenAI said and what the contract allowed

OpenAI stated that it could not be confident SpaceX would use its technology within the terms of service, citing its experience of Elon Musk's companies violating contracts. It proposed a shutoff date of 12 November 2026.

The mechanism is more mundane than the language. OpenAI's agreement with Cursor gave it a limited window to cancel following a change of control, and SpaceX closed its US$60bn acquisition of Anysphere earlier this month. The window was open and OpenAI used it.

The two justifications do different work. Enforcing a contract clause requires no argument about anyone's character, whereas the public statement of distrust is a reason offered rather than a right exercised.

12 NovemberProposed shutoff date
~5%Share of Cursor traffic on OpenAI models, per its chief executive
US$60bnWhat SpaceX paid for Anysphere
Change of controlThe clause that opened the cancellation window

Five per cent is the number that matters

Michael Truell, Cursor's chief executive, put OpenAI's share of traffic at around 5 per cent. If that figure is accurate, the operational impact on Cursor users is small, and the cut-off framing overstates it considerably.

It also tells you something about how Cursor is built. A coding tool routing 95 per cent of its work elsewhere has already made itself model-plural, which is the architecture every serious tool in this category has converged on. We covered the same design when Warp built its agent pipeline to be model-neutral.

The figure comes from Cursor and has not been independently verified. A company facing a public termination has an obvious interest in that number being small.

Anthropic gains and OpenAI knows it

Anthropic responded by promising expanded Claude support in Cursor. That was the predictable outcome, and for OpenAI it reads as a cost accepted rather than an advantage sought.

Removing your models from a widely used coding surface hands that surface's share to your closest competitor. Cursor's users do not stop coding; they change which model answers. If OpenAI's presence really was 5 per cent, the transfer is small. If it was more, OpenAI has paid a real price for the principle.

In the wider context, model access in coding tools stopped being a neutral utility some time ago and became a distribution channel worth fighting over. Anthropic became Slack's default model this week, and defaults in these surfaces decide usage more effectively than benchmarks do.

Why SpaceX bought a coding tool at all

The acquisition itself is worth a second look, because a rocket company buying a code editor for US$60bn is not self-explanatory.

The through-line is that Musk's companies already train on Cursor. Grok 4.5 was described at launch as SpaceXAI's first Cursor-trained flagship, which makes the editor a data source and a distribution surface for models the same group builds.

Owning Cursor converts a dependency into an asset. SpaceX now holds the interaction data, the tool's position in a developer's workflow, and the power to decide which models get called first. That is the same logic behind every other acquisition in this layer, and it is precisely why OpenAI would rather not be a supplier inside it.

That reframes the termination. When a supplier is feeding a competitor's product and has a contractual way out, the decision to leave is not necessarily about trust.

Change-of-control clauses are about to matter more

The transferable lesson here has nothing to do with Musk or Altman.

Almost every AI application is built on model access it does not own, under a commercial agreement it did not write. Those agreements routinely contain change-of-control provisions, because a supplier reasonably wants a say in who ends up as its counterparty.

The consequence is that an acquisition can strip a company of a core dependency at the supplier's discretion, within a window the acquirer does not control. That risk is rarely priced into a deal and is almost never visible to the acquired company's customers, who discover it when a shutoff date is announced.

For anyone building on a model API, the practical question is what happens to your agreement if your company is acquired, and whether you could route around the loss inside the notice period. Cursor appears able to. Most companies building on a single provider would not be.

What this does to the model-neutrality argument

Coding tools have spent two years telling customers that model choice is theirs. This is the first significant test of what that promise is worth when the choice is removed from above.

Model neutrality always had a hidden dependency. It requires every provider to remain a willing supplier. A tool can architect for plurality, route intelligently and expose a model picker, and none of that survives a supplier deciding it no longer wishes to be on the list.

Cursor's position is that the loss is 5 per cent, which is the best possible version of this outcome and is only available because it had already diversified. A tool that had standardised on one provider for quality reasons would be facing a rebuild rather than a routing change.

The lesson for buyers is that a model picker in the interface is not the same as supply security underneath it. Ask a vendor how many providers it holds commercial agreements with. That number, rather than the count in the model dropdown, is what survives a dispute like this one.

What to watch

Three things would clarify the size of this.

Whether the 12 November date holds, since it was described as proposed rather than fixed and these disputes frequently settle. Whether Cursor publishes anything about model mix afterwards, which would test the 5 per cent claim. And whether OpenAI applies the same reasoning to other SpaceX-owned or Musk-affiliated products, which would establish this as a policy rather than a single decision.

For a developer using Cursor the practical advice is unglamorous: check which model your tool is calling, and know what it falls back to. That is good practice regardless of this dispute, which serves as a useful reminder.