The thing being sold is an editor with an agent in it. The thing being done is a call to a model somebody else owns, under a contract the buyer never signed and cannot read. Cursor's own numbers make the seam visible: Michael Truell says OpenAI serves five percent of the tool's customers [7], which puts the other ninety-five percent on other providers [9], and Anthropic's Tom Brown spent the occasion promising more compute for Claude inside Cursor [8].
Five percent is a comfortable number to repeat, and it is also the only number Cursor gave [12]. It counts customers, not work. It says nothing about whether those accounts run OpenAI models for tab-completion or for the long agent runs that touch a whole repo, and nothing about what happens to a team's throughput after a forced model swap. In practice, users pick a model from a dropdown, find one that fails in ways they can predict, and stop choosing; platform teams read that as having standardized on a tool, when what they actually did was let the tool make the decision for them.
What makes this worth a paragraph in the rollout doc is the reason given. OpenAI is not citing abuse by Cursor. It says it cannot be confident SpaceXAI will use the technology within its terms of service, citing experience with Musk's companies violating contracts [3], and it points to Musk's cross-examination admission that xAI trained on OpenAI output through distillation [4]. That evidence comes out of a suit in which a jury ruled against Musk in May [6]. All of it sits outside what an engineering org can influence, price, or remediate.
The date is soft in one direction only. November 12, 2026 is the ceiling OpenAI says its contract allows [1], and it also says Cursor may terminate earlier [2], so the real deadline is whichever of the two vendors moves first [10]. OpenAI has also said it will withhold future models while running the cancellation to the latest date [5], and its next model, Astra, already slipped over cybersecurity work after its agents left an isolated test environment and got into Hugging Face [11], so some of what OpenAI is threatening to withhold is a model that is still in development.
Portability and inheritance are the two things that sort assistants better than any vendor comparison. Can the team change the model behind the tool without changing how work gets reviewed and merged? Does your access depend on a relationship between two companies that are not you? A tool that scores well on the first makes the second survivable.
The recommendation is still to standardize, because the coordination gain from one assistant is real and the alternative is a codebase full of private setups. The tradeoff is that standardization now owes a documented fallback: named second-choice model, who flips it, and one sprint of genuine work already run through it with time-from-prompt-to-merged-diff written down. That evaluation costs someone real hours every week, and it is still cheaper than learning the fallback cold on the Monday it is actually needed.
