OpenAI Cut Off Cursor's Model Access. Only 5% of Its Traffic Depended on OpenAI.

Model access is a supply chain, and supply chains have single points of failure. Cursor found one last week, when OpenAI notified SpaceX that it would stop serving models to the coding tool SpaceX had just acquired for $60 billion.

The proposed shutoff is November 12. The share of Cursor traffic affected is 5%.

Both numbers matter, but the second one is the one enterprise AI teams should be writing down.

Why OpenAI ended Cursor's model access in 2026

The stated reason is contractual trust, not economics. From OpenAI's announcement: "We are making this choice because we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts."

Two prior incidents sit behind that sentence. After Musk acquired Twitter, now folded into SpaceX, the company broke the terms of a data licensing agreement reported at roughly $2 million a year. And xAI, also now under SpaceX, admitted to violating OpenAI's terms of service.

OpenAI's Thibault Sottiaux reduced the position to four words: "It boils down to trust."

OpenAI says it gave the maximum notice its contract permits, and that it will not ship future models to Cursor during the wind-down. The acquisition itself closed on August 14, making it the largest purchase of a venture-backed startup on record. Fifteen days separated the close from the cutoff.

How Cursor, OpenAI, Anthropic, and SpaceX responded

Anthropic's response is the operationally interesting one. It committed additional compute within hours, which is what a supplier does when a competitor vacates measurable share overnight.

Why Cursor's multi-provider architecture held the damage to 5%

A tool carrying roughly $2.6 billion in annualized revenue lost one of the two most prominent model providers in the market, and the operational impact is 5% of traffic.

That outcome was architectural, not lucky. Cursor never allowed a single provider to become load-bearing. It runs Anthropic, Google, and Grok models alongside OpenAI's. On November 13, 95% of its traffic is unaffected.

Run the counterfactual. At 60% concentration, the same announcement is a company-threatening event, triggered by a dispute Cursor was not party to and had no vote in. The difference between a migration ticket and an existential problem was a routing decision made years earlier.

This is the third such cutoff in fourteen months. Anthropic cut off Windsurf in June 2025 with under five days of notice. Anthropic revoked OpenAI's own Claude access in August 2025 over a terms violation. None of the three were caused by the companies that absorbed them.

What the OpenAI-Cursor cutoff means for enterprise model strategy

At scale, provider selection stops being a procurement preference and becomes a systems problem.

The relevant metric is not how many models an organization uses. It is what percentage of a revenue-critical workflow would fail if its largest provider terminated on contractual notice. Calculated per workflow, not per organization, because aggregate figures conceal the exposure: a company spreading traffic across four providers can still run one critical workflow entirely on a single model.

Gergely Orosz reached the conclusion in June 2026, before this cutoff happened: "Trusting any single AI vendor seems like an increasingly high risk for any team or company."

We covered what SpaceX was actually buying when the $60 billion Cursor deal was announced. This is the first cost of that purchase landing on users who had no say in it. The pattern behind it now runs to three model cutoffs in fourteen months, and the architectural response is the same one that governs multi-agent orchestration in production: workflow logic stays independent of whichever provider serves it.

The question for enterprise teams is not which provider deserves trust. It is what share of a critical workflow currently runs on any single one.

Cursor's answer was 5%. That is why it is negotiating rather than rebuilding.