OpenAI said Friday it will stop supplying its models to Cursor after SpaceX closed a $60 billion acquisition of the AI coding startup on Aug. 14. The proposed cutoff is Nov. 12, 2026, and OpenAI said it will not give Cursor future models. A vendor relationship that had looked, to Cursor’s customers, like ordinary plumbing is being reclassified as a risk the lab will not carry.
The explanation was not a pricing dispute and not, on its face, a claim that Cursor’s editor had broken a rule. The lab wrote it “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 new owner is the problem OpenAI is willing to name. The date, the refusal of future models, and the appeal to prior contract fights are the rest of the decision.
A $60 billion coding startup
Cursor is an AI coding startup: software that lives in a developer’s workflow, proposes the next function, rewrites a file, and, in the version of this market that grew up after large language models became useful at software, often sends those requests to someone else’s model. For years that someone else included OpenAI. The product was a router as much as it was a destination. Users did not have to know which lab answered a given prompt. They had to feel that the answer was fast and good.
That is the business SpaceX bought. On Aug. 14 the acquisition closed at $60 billion. The figure is extraordinary for a developer tool and ordinary for this cycle, in which the companies that help engineers write software have become strategic assets for the companies that want to build the next wave of models, machines, and infrastructure. SpaceX is not a conventional software acquirer. It is Elon Musk’s launch company. As of mid-August it also owns a coding assistant that had been a customer of the laboratory Musk co-founded and later took to court.
Control is the whole argument. Until Aug. 14, Cursor could be described as an independent application company buying model access on commercial terms. After Aug. 14 it is a SpaceX subsidiary. The models would still be the same weights, served through the same APIs, used by the same people typing in the same editor. OpenAI’s position is that the legal person on the other side of the API is no longer one it will serve.
The proposed cutoff, Nov. 12, 2026, is months away. OpenAI is not yanking keys overnight. It is posting a date, refusing future models, and leaving a window in which Cursor can migrate traffic, finish conversations it says are already underway, or both. The more important half of the threat is the future. Coding tools live on the frontier. A product that can use only last season’s model while rivals get the next one is not using a utility. It is using a fading line.
“Neutral infrastructure” and a 5 percent share
Cursor CEO Michael Truell said OpenAI models are about 5% of Cursor traffic and that the companies are talking, calling OpenAI “neutral infrastructure” Cursor has used for years. Each clause is doing a job. The traffic share tells customers and employees not to treat Friday as an extinction event. The talks tell OpenAI there is still a door. The phrase “neutral infrastructure” is the theory of the market Cursor grew up in.
That theory says frontier labs are utilities. A coding assistant should be able to call whatever model is useful the way a cloud app calls object storage from more than one vendor. Neutral means the model maker does not punish the application maker for also being a customer of a rival, or for being acquired, or for belonging to a person the lab’s executives already feud with. Truell is saying Cursor treated OpenAI that way for years. Friday’s letter says the utility has decided it is not a utility where SpaceX is concerned.
Whether 5% is the right measure of dependence is a separate question from whether it is the number the chief executive put on the record. Coding traffic is not uniform. Some calls are cheap autocomplete. Some are long agentic loops that chew through expensive context. Some are the demonstrations that win enterprise contracts. A small share of requests can still be a large share of perceived quality if the best OpenAI model is the one users reach for on hard tasks. Truell did not break the 5% down that way. He offered the aggregate, and he offered it while saying the companies are still talking.
The talks matter because the cutoff is proposed, dated, and not yet executed. Nov. 12, 2026, is a deadline with a calendar attached. It is far enough out that a reversal, a narrowed exception, or a transition plan could still be negotiated. OpenAI’s additional statement that it will not give Cursor future models is harder to walk back. Even if today’s models keep running until mid-November, the next generation would be withheld. In a market that retrains its sense of “good” every few months, that is the clause that turns a supply dispute into a product problem.
Terms of service, and a history OpenAI says it already has
Terms of service are the dullest documents in the AI industry and, in this fight, the sharpest. OpenAI’s public explanation is not that Cursor’s product is a competitive insult. It is that the lab cannot be confident SpaceX will use the technology within those terms, and that the lack of confidence is based on experience with Elon Musk’s companies violating contracts.
The lab did not, in the language reported Friday, publish a new forensic account of which contract was broken, when, or by which Musk company. It treated the pattern as known. Anyone following the last two years of lawsuits, countersuits, and social-media broadsides between Musk and OpenAI will recognize the posture. OpenAI is saying it has done this before and will not do it again with the same counterparty under a new logo.
That is a different claim from a simple competitive cutoff. AI labs have spent the last two years drawing lines around who may use their models for what — rival labs, military customers, countries, and, increasingly, companies controlled by people they are already in conflict with. A terms-of-service rationale puts the decision in the compliance file rather than the strategy file. It also puts SpaceX, and by extension Musk, in the position of answering a charge about contract behavior rather than a charge about product overlap.
There is a reason labs care about downstream use. Frontier models can be distilled, fine-tuned, probed, or used as a silent teacher for a competitor’s system. They can be wired into agents that take actions the original vendor never reviewed. They can be run in environments the vendor cannot inspect. Terms of service are how a lab tries to keep a commercial API from becoming a training pipeline or a loophole. OpenAI is asserting that, once the customer is SpaceX, it cannot be confident those terms will hold. Confidence, in this telling, is the whole product.
Musk: “I couldn’t care less”
Musk posted Saturday, “I couldn’t care less,” then called Sam Altman and Greg Brockman untrustworthy and repeated his claim they “stole” a nonprofit. The sequence is the feud in miniature. First, indifference: Cursor does not need OpenAI, or Musk does not need OpenAI’s blessing, or both. Then the older indictment: the chief executive and the president of OpenAI cannot be trusted, and the lab they run was taken from the nonprofit structure Musk says he signed up for.
Altman and Brockman are not bystanders in this story. They are the people Musk has been litigating against, posting about, and competing with. Naming them, rather than “OpenAI” as a corporate person, keeps the fight personal. The “stole” a nonprofit line is the core of Musk’s political and legal case against the company’s for-profit conversion: that a charitable mission was turned into a capital structure he says he would not have agreed to. He lost the 2024 lawsuit. The courtroom is still the place that argument was tested.
Musk’s Saturday post does not, on its face, engage OpenAI’s terms-of-service claim. He did not offer a point-by-point defense of how SpaceX would use Cursor’s models. He dismissed the cutoff and restated the character case. “I couldn’t care less” is also a message to Cursor’s users and to the engineers who will have to reroute that 5%. The new owner is not pleading for continued access.
Indifference as a public posture is not the same as indifference as an operating plan. A $60 billion acquisition of an AI coding startup is a bet that software-writing tools matter to whatever SpaceX intends to build next. Withholding future OpenAI models is a bet, on OpenAI’s side, that those tools should not be built on its weights. Musk’s reply refuses to treat the second bet as a crisis. It does not make the first bet smaller.
Anthropic, Claude, and the Windsurf bruise
Anthropic, which rents compute from SpaceX, said it will increase Claude support in Cursor, drawing jabs over Anthropic’s earlier Windsurf cutoff. The sentence is crowded with ironies the industry did not have to invent.
Anthropic is OpenAI’s most direct rival in the market for careful, expensive, developer-facing models. Claude is already one of the answers Cursor can give when a user asks a coding agent to work. Increasing Claude support is the obvious commercial move: if OpenAI is leaving, fill the socket. It is also awkward, because Anthropic’s compute sits, at least in part, on SpaceX. The same corporate family that OpenAI will not supply is a landlord to the lab that says it will supply more.
Windsurf is the bruise. Anthropic earlier cut that coding startup off. The details of that cutoff are not in Friday’s announcement, but the jabs are. Critics of Anthropic’s Cursor pledge did not need a new theory of hypocrisy. They had a recent example of the same category of company — an AI coding assistant — losing access to Claude. Increasing support for Cursor after cutting off Windsurf is, to those critics, a reminder that a lab’s enthusiasm for coding assistants is a policy that can change when the customer changes.
None of that makes Anthropic’s Cursor pledge imaginary. The company said it will increase Claude support. In a world where OpenAI will not give Cursor future models, that pledge is the most concrete substitute on the table. It does not restore the OpenAI models. It does not resolve OpenAI’s contract complaint. It does tell Cursor’s users that one frontier lab is moving toward the product as another moves away.
The compute relationship adds a second complication. If Anthropic rents from SpaceX, then SpaceX is both Cursor’s owner and Anthropic’s supplier. OpenAI’s cutoff is an attempt to keep its models out of that cluster. Anthropic’s expansion is a decision to put more of its models in. The industry’s infrastructure map and its product map no longer match. A rocket company owns a coding editor, rents compute to a model lab, and is being denied API access by a third lab whose leaders include people Musk calls untrustworthy.
A feud that outlasted a lawsuit
The feud sits on Musk’s lost 2024 lawsuit over OpenAI’s for-profit conversion. That is the legal floor under every subsequent insult. Musk sued. He argued, in essence, that OpenAI had abandoned the nonprofit bargain. He lost. The conversion, the capital raises, and the corporate structure that let the lab take in vast sums survived the case. The personal conflict did not.
Losing in 2024 did not end Musk’s claim that Altman and Brockman “stole” a nonprofit. He repeated it on Saturday. It did not make OpenAI willing to treat his companies as ordinary API customers. Friday’s letter cites experience with those companies violating contracts. It did not keep SpaceX from buying Cursor for $60 billion on Aug. 14. It did not keep Anthropic from renting SpaceX compute or from promising more Claude in the editor SpaceX now owns.
What the lost lawsuit did do was remove the hope that a court would settle the relationship. After 2024, the parties have been left with the tools they still control: model access, acquisition capital, social-media statements, and the loyalty of developers who just want the next function to appear in the editor. OpenAI used the first of those tools on Friday. SpaceX had already used the second. Musk used the third on Saturday. Truell is using the fourth, counting traffic shares and calling the departing vendor infrastructure.
The Nov. 12, 2026, date now sits on calendars at OpenAI, at SpaceX, and inside every company that has built a coding agent on top of someone else’s model. The lesson is not subtle. Neutral infrastructure is a description of how a customer wishes to be treated. It is not a right. A lab that cannot be confident will stop supplying, will withhold future models, and will write down, in public, that the new owner’s name is the reason. Cursor says that is about 5% of traffic, that talks continue, and that the hole can be filled — including, Anthropic says, with more Claude. The rest of the market will watch whether 5% stays 5% once the models that have not been trained yet never arrive.