OpenAI hands Cursor developers 76 days to migrate off its models

Counted out, the longest available notice runs 76 days, from the August 28 announcement to the November 12 shutoff ^[20], against a supply relationship OpenAI itself dates at nearly four years ^[13], which puts the runway at roughly five percent of the working history ^[21]. A change-of-control clause is cheap to write into a supply agreement and almost never exercised, so its whole value sits in the option, and the option turned out to be worth more to OpenAI than the revenue sitting on the other side of it.

Inside Cursor the mechanical switch is close to trivial, since Anthropic, Google and Grok-linked first-party models all still route through the same product ^[12], so a team changes vendor with a settings change and a new purchase order. What does not re-point on a Friday afternoon is everything fitted around one supplier's behaviour: the prompt libraries, the evals, the reviewer habits calibrated to a particular model's failure modes. Neither publisher attempts to price that, and neither puts a number on what Cursor paid OpenAI ^[26], which is the figure that decides whether this is expensive discipline or a cheap channel trade, given that Cursor had become one of the largest buyers of third-party models anywhere ^[14].

What OpenAI is not doing, as a result, is holding its position in the busiest developer surface it did not build. Astra will not ship there ^[8], current models stay live until the date ^[9], and the transition path OpenAI names runs partly toward its own coding tools ^[18]. The stated reason is trust rather than price ^[5], resting on the Twitter contract breach ^[6] and on Musk's sworn admission that xAI had violated the same terms ^[7], with OpenAI adding that a partner of this size would normally need a tailored contract to keep usage inside agreed rules and manage safety at scale ^[23].

On the buyer's side the arithmetic is worth holding still for: about $60bn in stock for Anysphere in June ^[10], six months after xAI came into SpaceX at a combined $1.25trn mark ^[11], so Cursor cost roughly 4.8 percent of that merged valuation ^[22] to acquire what PYMNTS called the product xAI lacked, a widely adopted coding tool with a large base of paying professional engineers ^[19]. Part of what that 4.8 percent bought was neutrality, which co-founder Michael Truell, now a SpaceX executive, says Cursor maintained for years ^[15]. That is the asset OpenAI has just written down.

This is probably wrong, but I think the November date holds. Truell says the companies are talking ^[15] and OpenAI left today's models running ^[9], which is what a counterparty keeping a door open does; against that, Musk's public reply was that he could not care less and that Altman and Brockman are untrustworthy ^[16], and he is still appealing the 2024 suit he lost ^[17]. The settlement version, the version where the cutoff simply lands and other labs absorb the traffic, and the version where those labs read their own change-of-control clauses the same way, all price very differently for anyone standardising an engineering org on a router. My read is wrong if the paying base mostly does nothing by November 13 and ships at the same rate, which would say the model underneath the tool was a commodity input all along.