What OpenAI did
SpaceX closed its $60B all-stock acquisition of Anysphere, the company behind Cursor, on 14 August, folding it into a new SpaceXAI division. It is the largest acquisition of an AI software company on record, and it followed SpaceX's merger with xAI in February.
Two weeks later OpenAI ended the partnership. Cursor's direct access to OpenAI models stops on 12 November. The stated reason is that OpenAI cannot be confident the new owner will operate inside its terms of service.
Read on its own that is a story about a coding tool and an API key. Read against the last eighteen months it is the same move, run by the company that was on the other side of it.
Anthropic ran this play in 2025
On 6 May 2025 Bloomberg reported that OpenAI had agreed to buy Windsurf for around $3B, which would have been its largest acquisition. Inside a month, Anthropic cut nearly all of Windsurf's direct access to Claude, with less than a week of notice.
Anthropic's co-founder was direct about why. It would be odd, he said, to be selling Claude to OpenAI. The company framed it as capacity going to lasting partnerships.
Two months after that, in August 2025, Anthropic revoked OpenAI's own access to Claude.
The Windsurf deal never closed. Microsoft held rights to IP OpenAI acquired, Windsurf would not hand that over, and when exclusivity lapsed on 11 July Google paid $2.4B to hire the chief executive, a co-founder and a handful of engineers and license the technology.
Cognition took what was left. The company came apart in about seventy-two hours.
Cursor was the reason given
This is the part worth sitting with. When Anthropic explained why Windsurf was losing access, it pointed to the customers it was protecting capacity for, and Cursor was one of the names.
Fourteen months later Cursor is on the receiving end of the same decision, made by the company that was buying Windsurf when Anthropic made it. Every participant has now been on both ends.
Anthropic cut Windsurf to keep Claude out of OpenAI's hands, then cut OpenAI directly. OpenAI, having lost the deal that started it, is cutting the customer Anthropic protected.
Nobody in this story has a principled position on model access. They have a competitive position, and it moves.
Eleven weeks against six days
One difference is worth naming, because it is the only part of this that actually helps developers.
Windsurf got under a week. Cursor gets about eleven weeks and a public post explaining the reasoning.
Eleven weeks is enough to migrate a product. Six days is enough to break one.
OpenAI also handled it as a communications event, with a statement, a timeline and an acknowledgement that the people worst affected are developers who chose those models inside Cursor.
Anthropic sent notice and let the story form on its own. That difference is the lesson of watching what happened to the last company that ran this play.
The dependency nobody prices in
Any company building on a frontier model carries two dependencies it does not control. The first is the model.
That one is understood, it gets discussed in diligence, and the mitigations are known: route through more than one provider, keep an abstraction layer, hold your own evals so a swap is a decision rather than a crisis.
The second is the story about why the model went away, and almost nobody prepares for it. Windsurf lost that one. For several days the only version in circulation was that Windsurf had been cut off, which reads as a company that did something wrong.
The version where a competitor removed a dependency ahead of an acquisition arrived later and travelled less far.
The commercial damage from losing an API is recoverable in a quarter. The damage from being the company something was done to takes longer, because it sets how every subsequent story about you gets written.
What we would run in the first 48 hours
If a provider does this to you, the framing is decided inside two days and mostly inside the first six hours. Four things matter.
Your version goes out first, or within the hour. Whoever publishes first sets the frame and everyone else responds to it. A statement that arrives the next morning is a reply, and replies read as defence.
Lead with what users should do. The audience that matters is people with your product open right now, and their question is whether anything breaks and when.
Answer that in the first two lines. The competitive dynamics go underneath, if at all.
Say the timeline out loud. A specific date is the single most calming thing you can publish, and it is the detail that gets quoted. Vague reassurance gets ignored and then gets contradicted.
The people who carry your version already use the product. They are developers with audiences who have been building on you for a year, and they will put it in their own words, which is the only form of this that reads as true.
That relationship is built long before you need it, which is the entire argument for holding an audience when nothing is wrong.
None of this is available to a company that only talks to the market on launch day. The companies that survive a story like this are the ones with somewhere to put their side of it, and an audience already listening. That is worth more than the API key, and unlike the API key, nobody can revoke it.