OpenAI cut off Cursor after SpaceX bought it — here's what citing 'contract history' actually means

OpenAI terminated Cursor's API access yesterday, hours after SpaceX announced it had acquired the AI coding assistant. The official reason: Elon Musk's "documented history of breaking contracts."

Cursor co-founder Michael Truell posted on X that OpenAI models accounted for just 5% of Cursor's AI traffic. The company runs primarily on Anthropic's Claude Sonnet 4.5 and their own fine-tuned models. Truell framed the cutoff as a non-event.

That's the wrong read. The real story is what OpenAI's move reveals about API dependency risk, even at 5%.

What the 5% number means

Cursor has 2 million users. If OpenAI models handle 5% of requests, that's still ~100,000 users who hit OpenAI endpoints daily. Those users likely skew toward specific workloads — maybe long-context diffs, maybe multi-file refactors where GPT-4o-mini is cheaper than Claude. Losing that 5% means either migrating those flows to other models (cost spike, latency change) or degrading the experience for a subset of users.

The 5% also tells you Cursor was already hedging. They didn't build on OpenAI as the primary backend. They built a multi-provider system with OpenAI as one option among several. That hedge is the only reason they can shrug off the cutoff publicly.

Companies that run 80% on OpenAI don't have that option. When the API access dies, the product dies.

The contract-history justification

OpenAI's stated reason — Musk's history of breaking contracts — is both specific and vague. Musk has been in litigation with OpenAI since 2024 over claims that the company violated its founding charter. He's also publicly attacked Sam Altman on X for years. OpenAI's TOS gives them discretion to terminate any customer "for cause," and "association with a party in ongoing litigation" probably qualifies under that clause.

But the speed matters. Cursor announced the SpaceX acquisition in the morning. OpenAI cut access by afternoon. That's not a legal review; that's a policy already in place. OpenAI likely flagged Musk-affiliated entities as ineligible for API access months ago, and the Cursor acquisition triggered an automatic cutoff.

The precedent is what matters. If you're building on OpenAI and your company gets acquired by someone OpenAI doesn't like, your access can vanish same-day. That's not a terms-of-service edge case; it's a structural dependency risk.

What SpaceX gets from Cursor

SpaceX builds Starship flight software, Starlink satellite firmware, and ground-station control systems. All of that is C++, Rust, and Python at massive scale. Cursor is one of the best AI coding assistants for those stacks. The acquisition makes sense as an internal tooling play.

It also makes sense as a Musk negotiating lever. If Cursor grows to 5 million users and OpenAI's API ban holds, that's 5 million developers who experience OpenAI as "the model we can't use because of some CEO beef." Musk benefits from that perception, even if Cursor already runs mostly on Claude.

OpenAI benefits from cutting access early, before Cursor becomes large enough that the ban looks petty. At 2 million users, it's a footnote. At 10 million, it's a story.

The lesson for anyone building on third-party models

Cursor's 5% OpenAI traffic is a hedge, not an accident. They route requests to multiple providers because no single provider is reliable enough to bet the product on. That's the correct strategy in 2026, and it's going to be the correct strategy in 2027.

If you're building voice agents, build for ElevenLabs and Deepgram and Cartesia. If you're building code tools, build for Claude and GPT-4o and Gemini 2.5 Flash. If you're building RAG, build for Pinecone and Qdrant and pgvector. The hedge costs more upfront (multi-provider abstractions, testing matrix, fallback logic), but it's cheaper than losing 80% of your product when one provider cuts you off.

The Cursor cutoff is a reminder that API access is not a guaranteed right. It's a discretionary service, and the discretion flows one direction. If your product dies when one provider says no, your product is too fragile to survive.

Cursor survived because they already knew that.