OpenAI's Cursor cutoff reclassifies model access as a supply-chain dependency

The 5% figure is where this starts. Cursor's public reply says OpenAI models are about 5% of its traffic, and the same reply declines to treat the decision as final ^[7]. That figure is Cursor's own, and it is an aggregate over a base the company describes as millions of developers ^[12]. For it to describe your exposure, two things would have to hold: your model mix matches the house average, and request share tracks the share of work you cannot let regress. Neither is usually true. A release-gating review agent pinned to one model is one line of config and a disproportionate share of the value.

The bigger question is how much work this creates. OpenAI's post does not identify which models are affected, and does not say whether developers keep a path to OpenAI through their own accounts or API keys ^[20]. Those are different projects. If bring-your-own-key survives, the fix is billing plus a config edit. If it does not, you are re-selecting a model inside an editor whose repository context and configured workflows were built around the old one ^[14].

The trigger itself is harder to pin down. OpenAI told Reuters it was not confident SpaceX would use its technology "within our terms of service" ^[1], and latent.space reports the blogpost also cites "our experience with Elon Musk's companies violating contracts" ^[3]. Neither is a signal a customer can poll. Cursor said SpaceX closed the all-stock acquisition on August 14, a deal Axios valued at $60B against the $29.3B post-money mark from November 2025 ^[11][8], about 2.05 times the earlier number ^[18]. Supply changed roughly two weeks after ownership did ^[4].

The reporting does not agree on the day. runtimewire dates the announcement August 29, 2026, and notes OpenAI never stated the cutoff year ^[5]. Reuters-sourced coverage puts it on August 28 ^[2], and the Economic Times says Friday ^[1]. August 28, 2026 was a Friday ^[17]. Measured from August 29, the notice period is 75 days ^[19], one day longer if the Reuters date is the right one.

Why this option was available to OpenAI now and not last year is the mechanism worth keeping. According to latent.space, a year ago Cursor appeared in the GPT-5 launch video and a cutoff was a nonstarter because Claude models led coding by a wide margin ^[21]. Now GPT 5.6 reads as a credible alternative to the Claude 5 series and Cursor is promoting Grok 4.6 ^[10], with SpaceX supplying what Cursor calls the world's largest GPU fleet and first-party models added to the lineup ^[15]. Cutoffs get cheap once neither side is load-bearing for the other.

There is precedent. latent.space notes Anthropic did this to Windsurf while OpenAI was weighing an acquisition of it ^[9]. Two cases is enough to call it a procurement pattern. The OpenAI Startup Fund, for its part, was in Anysphere's $8M seed in 2022 ^[13]; the cheque outlasted the partnership.

In my context the right tradeoff is to treat the editor as replaceable at the model layer and to pay for that: a second provider exercised weekly against an eval set I own rather than one the tool ships, and a written answer from the vendor on which models it can still reach after November 12 ^[5]. It costs some throughput, but it buys independence from a trigger that lives in someone else's shareholder register.