OpenAI models leave Cursor on November 12: what to move, and where

On August 28, OpenAI published a short note saying it has told SpaceX it intends to wind down the contract that supplies OpenAI models to Cursor. The reason given is not commercial: "we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts." The note cites Twitter breaking OpenAI's contract after its acquisition and xAI's admission, under oath, that it violated OpenAI's terms. Both companies are now part of SpaceX, and so is Cursor, since the $60 billion deal closed on August 14.

The date comes from OpenAI's post on X: "Under our proposal, Cursor's direct access to our models would end on November 12." OpenAI describes that as the maximum notice its contract allows. Cursor's CEO Michael Truell replied that OpenAI models serve about 5% of Cursor user traffic and that Cursor is speaking with OpenAI to resolve it.

Two things in the note are easy to skim past and are the ones that change plans.

What the notice says, and what it does not

The models Cursor lists from OpenAI today are GPT-5.6 Luna, GPT-5.6 Sol and GPT-5.6 Terra. Those keep working through Cursor's own access until the shutoff date. The note adds a second condition that applies immediately: OpenAI will hold the cancellation to the latest date it can "while not providing future models." It names the next one, Astra. So a Cursor user will not see Astra in the model picker at any point, whatever happens to the November date.

The date itself is a proposal. Cursor says it is negotiating, and the coverage has no response from SpaceX yet. That is the one thing worth watching; it is not a reason to wait. The direction is set by OpenAI's stated reason, and that reason does not go away in a negotiation about timing.

Nothing in the note touches your own OpenAI account. It is the contract between the two companies that ends, which is exactly why the routes below survive it.

Three routes, and what each one costs you

Route	What you keep	What changes
Your own OpenAI API key in Cursor	The Cursor agent and chat, same editor, same rules	You pay OpenAI per token instead of Cursor per request. Tab completion keeps running on Cursor's models. Cursor's documentation states that its Zero Data Retention policy does not apply when you use your own keys.
OpenAI's Codex extension inside Cursor	OpenAI models on OpenAI's terms, still in the same window	A second agent with its own sign-in, its own sandbox and its own instruction files. Your .cursor/rules do not apply to it.
A different model in Cursor	Everything Cursor-native: rules, Tab, cloud agents, team settings	Claude, Gemini, Grok and Composer are all in the picker. Prompts tuned to Sol's verbosity or Luna's price point need retesting.

The first route is documented at cursor.com/docs/settings/api-keys: Cursor accepts keys for OpenAI, Anthropic, Google, Azure OpenAI and AWS Bedrock. The second is documented on OpenAI's side; the Codex IDE extension lists Cursor as a supported editor and installs from cursor:extension/openai.chatgpt. Neither route is guaranteed in writing beyond November 12, and neither is part of the contract being cancelled. If you want a commitment, ask for one now, while both companies are publicly saying they want to support affected developers.

The third route is the one most teams will take, and it is the one with the least visible cost: a prompt library that was written against one model's habits. If your agent instructions say "be terse" because Sol is not, or your batch job picked Luna because it was the cheap option in the picker, those decisions need a rerun against the replacement, not a find-and-replace.

The inventory to run this week

You do not need to migrate anything before November. You do need to know where the dependency is, and most teams cannot answer that from memory.

1. The model picker. Ask each developer which model they actually use in Cursor for agent work. Truell's 5% is Cursor-wide; on your team it is either 0% or it is the person who does the release branch.
2. Team admin settings. If you restricted the model list for your organisation, check whether an OpenAI model is the enforced default. A default that disappears becomes whatever Cursor falls back to.
3. Cloud agents and automations. Anything scheduled or triggered from a repository event runs unattended. Those are the jobs that will fail quietly on November 13.
4. Compliance paperwork. If you documented "Zero Data Retention via Cursor" for an OpenAI-backed workflow, the bring-your-own-key route changes that sentence, and the Codex extension route moves the data terms to OpenAI's. Either is fine; both need writing down.

For each hit, decide which of the three rows applies and try it once before you need it.

The dependency you just found

This is the second time in 2026 a model provider has ended access to a harness rather than the other way round. In January, Anthropic revoked xAI's access to Claude over its terms on training competitors; according to an internal xAI memo reported at the time, Claude stopped responding inside Cursor for xAI staff. That was aimed at one customer. This one is aimed at the harness's owner, and it lands on every user of one provider's models inside it.

The transferable part has nothing to do with Musk. A model choice inside an editor rests on two contracts: yours with the editor, and the editor's with the model provider. You can read the first. You cannot read the second, and it can be cancelled with the notice period written into it, which in this case was about 75 days. The inference provider behind a tool is a dependency in its own right, and it is worth knowing, per workflow, which one it is.

When we wrote about the acquisition in June, the open question was whether Cursor's model-agnostic approach would survive inside a company with its own models. Fifteen days after closing, one of the three outside providers has answered for itself. Anthropic and Google have not said anything, and Cursor's own Composer and Grok models are the ones it prices most aggressively. If you are choosing between Cursor and Claude Code or Codex for new work, "which company holds the model contract" now belongs on the list next to price and quality.