OpenAI is cutting Cursor off on 12 November. The 5% figure is true and it is the wrong number.

What happened

SpaceX’s acquisition of Anysphere, parent of Cursor, was announced in June at roughly $60 billion in stock and closed on 14 August 2026. Fourteen days later, on 28 August, OpenAI notified Cursor that it is ending the contract under which it supplies models, with 12 November 2026 proposed as the cutoff.

The mechanism is a change-of-control clause — the ordinary contract term that lets a party revisit an agreement when the counterparty’s ownership changes, on the logic that you signed with one company and now face another. Nothing exotic was invoked. What is unusual is who invoked it and why.

OpenAI’s stated position, per its statement to press: it cannot be confident that SpaceX will use OpenAI technology within OpenAI’s terms of service, based on its experience with Elon Musk’s companies violating contracts. It pointed specifically at Musk’s admission under oath, during the 2024 litigation, that xAI used OpenAI model outputs through distillation to train competing models — conduct those terms prohibit. Reporting also notes the earlier episode in which Musk cut off OpenAI’s paid licence to Twitter data after acquiring the platform.

Cursor CEO Michael Truell responded publicly on the Friday night: “OpenAI models serve about 5% of Cursor user traffic, and we’re speaking with the OpenAI team to resolve this.” He noted Cursor was one of OpenAI’s earliest customers and that it had trusted the platform to be neutral infrastructure. Anthropic co-founder Tom Brown called Cursor a longtime trusted partner and said Anthropic would increase the compute supporting Claude inside it.

Two details in the reported terms deserve more weight than they have received. First, 12 November is a proposed transition deadline, not a confirmed end date — access could end earlier. Second, the terms distinguish present from future: the OpenAI models Cursor uses today continue through the transition, and future OpenAI models will not be provided to Cursor at all.

Why 5% is the true number and the wrong one

Truell’s figure is almost certainly accurate, and it answers a question about token volume. The question a buyer needs answered is about product surface, and those diverge sharply here.

Bring-your-own-key is the obvious mitigation, and it is a partial one. A personal OpenAI key preserves local Chat and Agent in the Cursor desktop app. It does not restore:

The reason is structural rather than a policy anyone chose. Those features route through Cursor’s own infrastructure, which means the API call is made by Cursor as a customer of OpenAI — precisely the commercial relationship being terminated. A key you hold covers paths where your client talks to OpenAI directly. It cannot re-admit Cursor’s servers as a permitted customer.

That inverts the usual reading. The 5% of traffic is the part you can most easily replace, because interactive model selection is a dropdown. The parts you cannot replace with a credential are the ones Cursor spent 2026 pushing hardest — the cloud agents that wake themselves on a pull request, Slack thread or schedule, and the automation surfaces built on top of them. If a team pinned a GPT model ID into one of those pipelines, 12 November is not a preference change. It is an outage with a date on it.

The half of the terms nobody is quoting

Freezing today’s models and withholding tomorrow’s are very different commitments, and the second is the one that compounds.

If Cursor kept current OpenAI access indefinitely, this would be a footnote. If it loses access on 12 November and nothing else changes, it is a one-time migration affecting a twentieth of traffic. But the reported position is that no future OpenAI model reaches Cursor at all. On 2026’s release cadence — the Sol price cut and its promotional expiry, the Codex harness turning into an orchestration platform, a steady run of tier and pricing moves — that is roughly a significant OpenAI event a month that Cursor structurally cannot pass through while Copilot, Codex and Windsurf can.

So the gap does not stay at 5%. It starts near 5% and widens at whatever rate OpenAI ships. For a product whose central promise is pick any model, that is a genuine and permanent reduction in the thing being sold, and it belongs in the renewal conversation rather than in the news cycle.

Model supply as a competitive instrument

The uncomfortable structural fact underneath this is that four companies supply nearly all frontier capability, and every tool built on top of them holds a contract that can be terminated.

Cursor is not defenceless — it retains Anthropic, Google, xAI and its own Composer line, promoted alongside Grok 4.6, and under SpaceX it has access to Colossus capacity to train more of its own. That is arguably the strategic logic of the acquisition: Cursor was paying retail to Anthropic and OpenAI, two direct competitors, on every request, while its new owner had idle training capacity. Losing one supplier accelerates a first-party push that was already underway.

Anthropic’s position is the more revealing one. Its terms bar customers from building products that compete with its services — broader language than OpenAI’s, and on its face sufficient to justify the same cutoff. Instead Anthropic said it would add compute. The reason is visible in its own supply chain: Anthropic has been buying compute from SpaceX, including Colossus capacity tied directly to higher Claude Code and API limits, an entanglement its S-1 disclosures made concrete. Enforcing terms against Cursor would mean enforcing them against its own landlord, and Cursor is also one of the largest channels through which Anthropic reaches working developers.

That is the honest shape of the market. Whether your tool keeps its models is decided by the supplier’s commercial position, not by your subscription — and the answer varies by supplier for reasons that have nothing to do with you.

What to do this week

1. Find out which model your work actually runs on. Auto mode, Claude, Gemini, Grok or Composer: nothing changes for you on 12 November. Deliberately selecting GPT: pick a replacement now and validate it over a week on your own repository and real tickets, not a benchmark table — you chose that model for a behaviour you observed, so verify the replacement on the same work.
2. Audit pinned model IDs in automation first. Cloud agents, Automations, CLI and SDK are the surfaces bring-your-own-key cannot reach. Grep your configs for OpenAI model strings today. This is the only part of this story with genuine outage risk, and it is invisible from the editor.
3. Do not treat a personal API key as a continuity plan. It restores interactive editing and nothing automated. If you were going to write “BYOK covers it” in a risk register, that sentence is wrong in the specific way that matters.
4. Re-price optionality at renewal, not today. Cursor keeps three of four major suppliers plus its own models, and the immediate capability loss for most teams is near zero. What changed is the trajectory — a frozen and then removed OpenAI shelf, against a first-party model house with an owner supplying the training compute. Compare against Claude Code and the rest of the field in the best AI coding tools guide, and if you are weighing the two directly, Cursor vs Claude Code is the head-to-head.
5. Run the change-of-control test on every AI tool you depend on. Three questions: which advertised models are supplied by a third party rather than owned by the vendor; which of those contracts survive if this vendor is acquired by a supplier’s rival; and which product features break outright when a supply contract ends, versus merely switching which model answers. Routing through a neutral gateway helps at the API layer and does nothing for a product surface you do not control. Write the answers down before you need them.

Honest caveats

This is not final. OpenAI’s notice is a proposed termination with a proposed date, Truell says talks are continuing, and both parties have commercial reasons to prefer a settlement to a public divorce. The 12 November date could move in either direction, including earlier.

The 5% figure comes from Cursor. It is the vendor’s own characterisation of its exposure, published while it is negotiating, and no independent measurement of Cursor’s model traffic mix exists. It is plausible and consistent with Cursor’s promotion of Composer and Grok, but treat it as a party’s statement rather than a verified metric.

Nothing here rests on this site testing the product. The feature-by-feature breakdown of what a personal API key does and does not restore is drawn from reporting on OpenAI’s and Cursor’s published guidance, not from an evaluation run here.

Update (1 September 2026): what the other version of this looks like. The argument above is that model neutrality inherited from a tool vendor is a contract term rather than an architecture. On 31 August the Pentagon shipped the counterexample: ChatGPT Mil and Grok for Government went live on GenAI.mil alongside Gemini, three frontier vendors behind one portal for 1.7 million users, with the department holding each supplier relationship directly. That is portability nobody can revoke by invoking a clause against an intermediary — and the same piece covers its uncomfortable second edge, which is that a buyer with real substitution power can route around any vendor term it dislikes, including the safety ones.

The Musk-specific reasoning is genuinely specific. OpenAI cited evidenced prior conduct, not a general policy against supplying competitors’ subsidiaries. Do not generalise this into “frontier labs will now cut off rivals’ tools.” Generalise the narrower and more useful point: the clause that made it possible is in almost every supply agreement, and it is available to any supplier that decides it does not trust who it is now dealing with.

Questions or corrections? Email Pick Right. Want the full list? See all news.