Termination notice targets Cursor's contract for OpenAI models

The developer who feels this is the one who wrote a prompt library, a small eval suite and two tool-calling workarounds against one GPT model a year ago, shipped it, and has not opened that config since, rather than the one who casually tries three models on a hard bug. Cursor sells a picker with several vendors in it ^[3]. What people do with a picker is choose once, get an acceptable answer, and stop touching it.

So the traffic number and the migration number measure different things. The load side is the smaller problem. If OpenAI models carry about 5% of Cursor's traffic ^[1], everything else carries 95%, and absorbing every displaced request raises the load on the remainder by roughly 5.3% ^[4]. That is a capacity question a vendor can answer in a sentence, which is close to what Anthropic's Tom Brown did ^[2]. The rewrite side is a person-weeks question, and it sits with whoever tuned the prompts.

The Claude Code allowance deserves arithmetic rather than applause. Take a baseline weekly limit of 100 units: the boost in place now puts a user at 150, and the permanent increase puts them at 125 ^[9], which is about 17% below the ceiling they have today ^[5]. "Permanent" is doing real work in that sentence, and it is the word a capacity planner should read twice.

The report is also unsettled about who sent the letter. Digital Trends frames the story as OpenAI pulling its models and sources the account to OpenAI, while the sentence carrying the November 12, 2026 date names SpaceX as the party notifying Cursor of intent to terminate ^[8][6]. The same piece places the episode in the deteriorating relationship between Elon Musk and Sam Altman, and notes Musk criticised Altman after the news ^[10]. Until the notice itself surfaces, the date is the part worth putting in a plan and the counterparty is the part worth hedging.

The sort that helps here has two axes: share of traffic on one, and how pinned the workflow is on the other, where pinned means prompts, evals and output parsers tuned to one model's habits. Low share and unpinned is a dropdown change. High share and unpinned is fine as long as someone has actually run the swap. Low share and pinned is where Cursor's affected users are sitting, and it is the quadrant teams under-fund because the traffic percentage looks harmless in a review deck. High share and pinned is the one that belongs in a contract, with a notice period you can name.

The forcing question is narrow: a 30-day termination letter from either model provider tomorrow would cost some number of engineer-days to reach acceptable output elsewhere, and land on someone's desk to absorb. A vendor promising more compute has not answered either count.