OpenAI Pulled Its Models Out of Cursor. The Real Lesson Isn't About Musk.

Cursor has about four million developers. On 29 August, OpenAI told them GPT models would stop working in their editor. That was two weeks after SpaceX closed its $60 billion purchase of the company.

The Hacker News thread ran to 852 points and 546 comments, mostly about hypocrisy and Elon Musk. Both are fair. Neither is the real lesson. The lesson is that the model in your editor is a supplier deal, and a supplier can end it with one blog post.

Two weeks from record deal to cutoff

Here's the timeline, because the speed is half the point.

For scale: Cursor passed $4 billion in yearly recurring revenue in June, so SpaceX paid about 15 times revenue. Pro costs $20 a month and Business is $40 a seat. Those four million developers are the asset, and OpenAI just took a piece out of their model access.

What OpenAI said

The post is short. The line everyone quoted: "We cannot be confident that SpaceX will use our technology within our terms of service."

The background makes it less abstract. xAI has admitted under oath that it partly trained on OpenAI's models by distilling them, which OpenAI says breaks its terms. Anthropic had already blocked xAI from Claude months earlier for the same reason, and xAI engineers reportedly kept the pipeline running for a while through personal accounts and an encrypted relay. So OpenAI's point is simple. The company that now owns Cursor has already used our output to train a rival, and Cursor is a firehose of our output.

It also helps to remember that OpenAI owns Windsurf, a direct Cursor rival. The move protects a product as much as a principle.

Why Anthropic stayed, and why that isn't principle either

Anthropic went the other way. Co-founder Tom Brown said Anthropic will keep giving Claude to Cursor and add more compute for it. At the same time, Anthropic is renting a lot of compute from SpaceX.

The Hacker News read was blunt: the compute deal beats the policy. I think that's right, and I'd go further. Neither choice was about ethics. OpenAI has a rival to starve and a product to protect. Anthropic has GPUs to buy and a customer with four million seats. Both did the math, and the math came out different.

That's not a complaint. That's the point. When a supplier's interests change, your access changes with them, and nothing in your subscription says otherwise.

The part of the thread that was right

Three comments deserve more than a scroll past.

The top one made a point about double standards. The labs trained on the whole web and argued in court that it was fair use. But a model learning from another model's output is a terms violation. If a book can be turned into a model, the commenter asked, why isn't a model learning from a model just as fair?

A second pointed at companies. Plenty of them standardised on Cursor before anyone heard of the SpaceX deal. They now get Grok as the default and lose GPT, with no vote and no time to move.

A third pointed at who gets punished. Commenters noted that DeepSeek and Moonshot are widely thought to have distilled the same Western models, and nobody cut them off. The rule gets used when it's convenient.

The other Cursor story: it was already being used as a weapon

There's a second Cursor story from the same two weeks that almost nobody linked to the first. Reuters reported that a Russian-speaking ransomware group called Aurora used Cursor to attack at least seven companies earlier this year, including Christeyns in Belgium, Teckentrup in Germany, and Bayou Title in Louisiana.

The method matters more than the victims. The attackers told the agent the break-ins were authorised tests and asked it to hunt for credentials. Cursor said no. So they started a new session and asked again, and kept doing that until it said yes. Researchers found more than two dozen of these chats. The model underneath was Claude Sonnet 4.5.

If you build anything on top of an agent, that's the lesson. A refusal that resets when the session resets is a speed bump, not a wall. Real limits live outside the chat: in what the tool can reach, what it can touch, and what gets logged.

How I plan for the dependency

I'm not telling anyone to leave Cursor. What I'd change is how you treat it, in any editor.

Keep your rules, prompts, and skills in the repo, not in the editor's settings. If you have to open a different tool tomorrow, the setup comes with you. Keep the workflow working on at least two model vendors, so losing one is an annoyance, not an outage. Treat the editor's agent like any outside service: know what it sends where, and know how to turn it off. And don't let your team's whole AI setup live in one editor, because an editor is now a thing that can be bought.

The other tools are real: VS Code with extensions, Zed, Claude Code, GitHub Copilot, and Windsurf, which OpenAI owns. But switching isn't the point. Being able to switch is. Pick what you like, and make sure you could leave it in an afternoon.

Questions I keep seeing

Can I still use GPT models in Cursor?

Not through Cursor's own model access. OpenAI said it's ending that after the SpaceX deal closed. Anything past that depends on OpenAI's terms and can change, so check the current state before you rely on it.

Did Anthropic block Cursor?

No. Anthropic said it will keep Claude in Cursor and add capacity. It had blocked xAI earlier for distilling its models, so the two calls look inconsistent, and the compute deal with SpaceX is the likely reason.

Should I leave Cursor?

Only if the product stops working for you. The better move is to make leaving cheap: keep your rules and prompts in the repo, keep a second vendor working, and know exactly what the agent can reach.

OpenAI's post is here, and the Hacker News thread is worth the scroll. If you want the longer argument for why the base of your workflow should be yours and not the model's, it's in an earlier post.