OpenAI Cuts Off Cursor's Access to Its Models
OpenAI gave SpaceX notice on 28 August that it is ending the contract behind OpenAI models in Cursor. The proposed shutoff is 12 November 2026.
OpenAI has told SpaceX it will end the contract that puts OpenAI models inside Cursor, with a proposed shutoff date of 12 November 2026. The notice went out on 28 August, fourteen days after SpaceX closed its $60 billion all-stock acquisition of Anysphere, Cursor's parent company. The stated reason is not performance, price or payment. OpenAI says it cannot be confident SpaceX will use its technology within the terms of service, and points to previous contract disputes involving Musk-owned companies.
If you write code in Cursor, nothing breaks today. You have roughly ten weeks, and the practical question is narrower than the headlines suggest.
The timeline, in dates
14 August 2026: SpaceX completes the all-stock purchase of Anysphere, exchanging roughly 389.3 million SpaceX Class A shares.
28 August 2026: OpenAI notifies SpaceX that it intends to wind down the agreement, giving the maximum notice its contract allows.
12 November 2026: the proposed date on which OpenAI models stop being served through Cursor.
OpenAI published its reasoning in a note on its own site, and the decision was reported the same week by CNBC.
The clause that made this possible
The interesting detail is not the falling out. It is that OpenAI did not have to argue a breach to get out. Its agreement with Cursor carried an ownership-change termination provision, the kind of clause that lets either side walk when the counterparty is acquired. SpaceX buying Anysphere triggered it.
That matters well beyond this one deal. Ownership-change clauses are common in model-supply contracts and almost never read by the developers who depend on them. The vendor you signed with is not necessarily the vendor you will have next year, and the terms can move under you without anyone doing anything wrong.
Why 5% is the number to hold on to
Cursor co-founder Michael Truell put OpenAI models at around 5% of Cursor's traffic. Take that at face value and the blast radius is small: most Cursor users are already on Anthropic or Google models by default and will not notice 12 November at all.
The 5% is not evenly spread, though. It concentrates in people who deliberately pinned an OpenAI model, usually for a specific reason: a task where they benchmarked it as better, a prompt library tuned to its quirks, or a workflow that predates the alternatives being good. Those users absorb the entire cost of this change.
Three things to check before November
Find out what you are actually running. Open your Cursor model settings and note the exact model behind each mode. If nothing there starts with a GPT identifier, you are done.
Check pinned models in config files, not just the UI. Team-level rules files and per-project settings can pin a model that the interface does not surface.
Test a replacement now, while you still have both. Run the same three or four representative tasks through your current model and a candidate, and compare the diffs rather than the vibes. This is the same playbook you would use for a deprecation, and it is much easier to do while the old option still works.
What this does not change
Cursor is not losing its ability to run frontier models. Anthropic and Google models remain available in the product, and the tool itself is unaffected. This is a supply-side dispute between two companies, not a judgement about Cursor's engineering.
It is also not a reason to switch tools in a hurry. If your editor works and your models are not the ones being withdrawn, the correct response is to note the date and move on. If you are weighing options anyway, how the main coding tools compare is a better starting point than a news cycle.
The pattern worth noticing
Model access is turning into a commercial lever rather than a utility. A year ago the assumption was that every serious model would be reachable from every serious tool, priced by the token. That assumption is thinning, and the people most exposed are the ones whose workflow depends on one specific model inside one specific product.
The cheap insurance is knowing which model you are on and having tested one alternative. That takes an afternoon, and it is worth doing before the next notice lands rather than after. If you would rather not track this yourself, keeping up with AI news without reading everything covers the filter.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
