OpenAI Cut Off Cursor. The Workaround Covers Less Than the Product Does.

On November 12, three of the eleven models inside Cursor stop working, and the fix its maker is offering restores fewer of them than most coverage implies. You can plug in your own OpenAI key, but only for local chat and agent requests. Autocomplete, automatic model routing, background agents, automations and the command line stay dark.

That gap between “there’s a workaround” and what the workaround covers is the part worth your time, and it has nothing to do with code editors. If you buy any AI tool on a monthly plan, you’re in the same position: you have a contract with the tool, the tool has a separate contract with whoever makes the model, and you can’t see the second one. When it ends, you find out how far your vendor’s Plan B reaches.

Our reader mostly doesn’t use Cursor. Read it as the first clean public example of a marketing-relevant failure mode, because the same structure sits under your writing assistant, your ad-copy generator and the AI features inside your CRM.

What happened

OpenAI announced on August 28 that it’s ending the agreement putting its models inside Cursor, effective November 12. The trigger was ownership: SpaceX signed a $60 billion all-stock deal for Cursor’s parent Anysphere on June 16 and closed it on August 14, folding the company into a new SpaceXAI division.

OpenAI’s contract carried a change-of-control right, which is a clause letting one side exit when the other gets bought. It used it, and said so bluntly:

“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 76 days between announcement and shutoff were not a courtesy. OpenAI described that window as the maximum notice provided by its contract. It gave everything it was obliged to give and nothing more.

The number that keeps this honest

Cursor co-founder and chief executive Michael Truell put the practical damage at about 5% of Cursor user traffic. Three affected models out of eleven in the lineup. Grok, Claude and Gemini all stay.

Take that number seriously, because it’s the one that stops this being a disaster story. Almost nobody loses their tool here. Truell’s own framing is what makes it interesting instead:

“Cursor was one of the very first users of OpenAI, we’ve worked closely with their team for years, and we’ve trusted their platform to be neutral infrastructure.”

The word doing the work is “infrastructure.” Infrastructure is a thing you build on because it doesn’t have opinions about you. What Cursor had was a commercial agreement with a competitor, containing a clause about who owns them. Those are different objects, and the 24/7 Wall St. account of the fallout makes clear that the exit had nothing to do with how Cursor behaved.

Why the fallback is thinner than it sounds

Here’s the detail nearly every writeup dropped. Bringing your own OpenAI key sounds like a complete answer, and it covers local chat and agent requests only.

It doesn’t restore autocomplete as you type. It doesn’t restore automatic routing, the feature that picks a model for you. It doesn’t restore background agents, automations, the command line or the programming interface.

Those are the surfaces where the tool does work without you watching. Which means the fallback preserves the part you drive by hand and drops the part that runs on its own, and the second one is why anybody pays for these tools.

Translate that to a marketing stack. If your content platform loses a model supplier tomorrow and tells you to bring your own key, ask which features that key lights up. The answer will likely be the chat box, not the scheduled brief generation, not the bulk rewrite, not the automated first draft that lands before you open the tab.

The version of this that reaches marketing

Nothing about a code editor makes this specific to engineering. Three things generalise.

Your vendor’s model supply is a contract with an exit. Not a feature, not a pipe. It has parties, terms and a clause about ownership. You signed a different contract, with someone who has one of these.

Change of control is the trigger nobody prices. Cursor did not lose access for anything it did. It lost access because it got acquired, and the acquirer had history with the supplier. Your vendor being bought is normally a footnote in your renewal, and it just became a reason a feature disappears.

You can’t audit any of this from the outside. Nothing in Cursor’s pricing page listed which models sat behind which features. It took a shutoff notice to reveal that autocomplete and background jobs were on a different footing from the chat box. This is the same visibility problem as AI assistants reaching into ad accounts without IT seeing it: the dependency runs the product, and nobody writes it down until it breaks.

David Frankel’s test from our piece on which martech survives sorts tools by what breaks if they disappear. This adds a second axis. A tool can be perfectly embedded in your workflow and still be sitting on a supply agreement that a corporate action ends.

The counter-case

Two things cut against reading this too widely.

The first is the 5%. This is a small operational event dressed in a large corporate feud, and treating every vendor relationship as fragile because of one clause between two companies with a public history would be an overcorrection. Most model supply is boring and stays boring.

The second is that switching worked. Cursor kept Grok, Claude and Gemini, and Anthropic moved quickly to expand support. The model layer turned out to be more substitutable than the alarm suggested, which is an argument that this kind of event is survivable rather than one that it’s coming for everyone.

Both are fair. Neither tells you which of your own features would go dark, which is the only question you can act on.

What to ask before your next renewal

One question, and it takes an email.

Ask each AI vendor you pay: which models power which features, and what happens to each feature if a model provider ends the agreement? Not “do you have redundancy,” which gets a yes from everyone. Feature by feature, because that’s the level at which Cursor’s answer turned out to be uneven.

A vendor who can answer that has thought about it. A vendor who can’t just told you the same thing Cursor’s users learned on August 28, only earlier and for free.