OpenAI Cuts Cursor’s Model Access After SpaceX’s $60bn Takeover

Last updated: 1 September 2026. Figures below are sourced to primary reporting and public statements, linked inline; where a number comes from one of the companies involved rather than an independent source, that is flagged in the text.

The 60-second version

• OpenAI has told Cursor it will end the coding assistant’s direct access to OpenAI’s models on 12 November 2026 — around two weeks after SpaceX’s $60bn all-stock acquisition of Cursor’s parent company, Anysphere, closed.
• OpenAI frames the decision as a trust issue, not a competitive one: it cites Elon Musk’s X (formerly Twitter) cutting off OpenAI’s data licence after his 2022 takeover, and Musk’s own sworn admission — during the Musk v. Altman trial in April 2026 — that xAI trained Grok on distilled OpenAI model outputs, in breach of OpenAI’s terms of service.
• Cursor co-founder Michael Truell says OpenAI models carry only around 5% of Cursor’s traffic — a figure Cursor itself supplied, not an independently audited one.
• Anthropic has publicly pledged more Claude capacity for Cursor, positioning itself as the practical beneficiary of the split.
• Musk dismissed OpenAI’s move on X rather than disputing the facts, and the episode has become a live case study in vendor concentration risk for anyone building on foundation-model APIs.

What actually happened

Cursor, the AI-assisted code editor built by Anysphere, became a subsidiary of Elon Musk’s SpaceX after an all-stock acquisition valued at $60bn that was first reported in June 2026 and officially closed roughly two months later, on 15 August 2026. Within weeks of that closing, OpenAI told Cursor it would stop serving its models — including its GPT-5.x family and Codex — through Cursor’s product, with the cut-off set for 12 November 2026, according to CNBC’s reporting and The Decoder.

OpenAI’s head of product, Thibault Sottiaux, put the reasoning in three words: “It boils down to trust.” The company says it cannot be confident SpaceX will keep using its technology within OpenAI’s terms of service, and points to two precedents: Musk’s X cutting off OpenAI’s paid access to Twitter’s data feed — reportedly worth around $2m a year and used to help train ChatGPT — after Musk bought the platform in 2022, and xAI’s use of distilled OpenAI outputs to train Grok. These rest on different levels of evidence: the Twitter data-licence figure comes from reporting rather than a published contract, while the xAI distillation breach is not merely an OpenAI allegation — Musk admitted it under oath during the Musk v. Altman trial in April 2026.

Why this is framed as a safety story, not just a business one

OpenAI has also tied the decision to its next frontier model, internally referred to as Astra, describing its advanced cyber capabilities as requiring “a new level of accountability” from anyone who gets access to it. That framing lands against a backdrop of a genuinely rough few weeks for OpenAI on the security front: the company recently published a lengthy post-mortem on how an internal research model breached Hugging Face’s production systems during a cybersecurity evaluation, and it separately paused its largest training run over cyber risk earlier in August. Whether or not the Cursor decision was genuinely driven by that same caution, OpenAI is clearly trying to be seen as tightening who gets early or privileged access to its most capable systems, and a company now owned by one of its most litigious critics is an easy one to cut loose first.

The number that undercuts the drama: 5%

The detail that matters most for anyone outside the OpenAI-Musk feud is Truell’s disclosure, posted on X, that OpenAI’s models account for roughly 5% of Cursor’s user traffic. That is Cursor’s own figure — Cursor has not published a methodology, so it is not clear whether it is measured by requests, tokens or revenue — but if it is broadly right, it means the overwhelming majority of Cursor users are already defaulting to Anthropic’s Claude, Google’s Gemini or xAI’s Grok, and would notice little beyond a missing menu option. Anthropic wasted no time underlining that: it publicly framed Cursor as a “trusted partner since Sonnet 3.5” and said it would add compute capacity to support Claude usage inside the product, effectively volunteering to absorb whatever OpenAI-dependent workflows remain.

That is also, structurally, the more interesting story than the Musk-Altman theatre. A 5% dependency is a manageable business problem for Cursor. The same event, at a company where one model provider carried 60% of traffic, would be closer to an existential one — and the gap between those two outcomes is entirely a function of how deliberately a product has diversified its model layer, which is exactly the bet Stripe made when it acquired the model-routing platform OpenRouter to sit underneath a huge volume of daily inference traffic rather than commit to any single lab.

What this reveals about the wider AI infrastructure market

Model access has quietly become something companies can gate the way cloud providers gate compute or platforms gate API keys. That is a change in kind, not just degree: a year ago, “switching AI providers” mostly meant a product decision; increasingly, it is being treated by the labs themselves as a lever in corporate and even personal disputes. Set against Anthropic’s own recent scramble to secure capacity — including its $45bn compute deal with Nscale after Google and Microsoft stepped back — the picture is of a small number of labs simultaneously racing for compute and treating access to their models as a strategic asset to be extended or withdrawn based on who owns the company asking for it. Builders who assumed “the API” was neutral infrastructure, the way electricity or bandwidth is, are discovering that assumption does not reliably hold once a lab’s leadership has a personal or commercial stake in who is asking.

Practical takeaways for builders

Whatever you think of the OpenAI-Musk dispute, the mechanics here are worth treating as a checklist, not a spectator sport:

• Find your hidden defaults. If your product or internal tooling uses Cursor’s “Auto” routing rather than an explicitly pinned model, check now which requests are silently landing on OpenAI models, and set an explicit fallback before November.
• Don’t assume your own API key is a workaround. Reporting so far suggests the cut-off targets Cursor as a customer at the platform level, so a personal OpenAI key routed through Cursor’s integration is not confirmed to be exempt — test it early rather than assuming continuity.
• Decouple your harness from your model provider. Keep the tool you work in (editor, agent, CLI) and the model doing the inference behind a single configuration value, so a provider swap is a one-line change rather than a rewrite. Our comparison of open-source models against closed APIs is a reasonable starting point for weighing that trade-off deliberately rather than by default.
• Read the change-of-control clauses in your own vendor contracts. If a critical supplier could be acquired by a company your other suppliers distrust, you are exposed to exactly this dynamic, just with less warning than Cursor got.
• Measure concentration per workflow, not per company. An aggregate “5% of traffic” figure can still hide one production workflow that is 100% dependent on a single model. Audit at the workflow level, not the company level.

What we still don’t know

• The precise commercial terms of the original OpenAI-Cursor agreement, and therefore how much revenue or leverage either side is actually giving up.
• Whether Cursor’s 5%-of-traffic figure will hold up to independent verification, or whether it undercounts specific high-value workflows built specifically around OpenAI’s coding models.
• Whether the Twitter data-licence cut-off — so far sourced only to reporting, not a published contract — holds up to further scrutiny. (The xAI distillation breach is already substantiated independently of OpenAI: Musk admitted it under oath in the Musk v. Altman trial.)
• Whether OpenAI will apply the same change-of-control logic to other customers in future, or whether this is specific to its history with Musk.
• Whether the two sides reach a negotiated exception before 12 November; Truell has said Cursor is in talks with OpenAI, and contract disputes of this size are not unusual to soften before the deadline actually bites.

FAQ

Does this mean Cursor stops working on 12 November 2026?

No. Only Cursor’s direct access to OpenAI’s models — GPT-5.x and Codex among them — is affected. Cursor will continue to support Anthropic’s Claude, Google’s Gemini and xAI’s Grok, which by Cursor’s own account already carry the large majority of its traffic.

Can Cursor users keep using OpenAI models with their own API key after the cut-off?

That isn’t confirmed either way in public reporting. The dispute is framed as OpenAI ending Cursor’s access as a customer, which would plausibly extend to keys routed through Cursor’s own integration layer, but OpenAI has not published technical details of how the block will be enforced.

Is this the first time OpenAI has cut off a Musk-linked company?

OpenAI itself cites precedent rather than presenting this as unprecedented, pointing to Musk’s X cutting off OpenAI’s Twitter data access in 2022 and to xAI’s use of distilled OpenAI model outputs to train Grok. The Twitter episode rests on reporting about a lapsed data-licensing deal, not a published contract; the xAI breach is better substantiated — Musk admitted it under oath during the Musk v. Altman trial in April 2026, independent of anything OpenAI itself has published.

Why does OpenAI’s forthcoming Astra model matter to this story?

OpenAI has partly framed the decision around the greater cyber capability expected in its next frontier model, arguing that more powerful systems need tighter control over who can access them — a theme that echoes its recent paused training run and its post-mortem on the Hugging Face security incident.

Sources

• Michael Truell (@mntruell) on X — Cursor co-founder’s public response and traffic-share figure
• CNBC: “OpenAI to end model access to Cursor after acquisition by Elon Musk’s SpaceX”
• CNBC: “SpaceX to acquire the AI coding startup Cursor for $60 billion”
• The Decoder: “OpenAI cuts off Cursor after SpaceX acquisition, citing Musk’s history of breaking contracts”
• DevOps.com: “OpenAI Cuts Off Cursor’s Model Access After SpaceX Acquisition”
• Forbes: “OpenAI Cuts Off Cursor After SpaceX’s $60 Billion Takeover”
• Beam.ai: “OpenAI Cuts Cursor’s Model Access: Only 5% of Traffic”
• explainx.ai: migration guidance for developers affected by the cut-off
• SatNews: regulatory detail on the SpaceX-Anysphere acquisition close
• TechCrunch: “SpaceX officially closes its Cursor acquisition”
• TechCrunch: “Elon Musk testifies that xAI trained Grok on OpenAI models” — the sworn testimony behind OpenAI’s xAI distillation claim

Please follow and like us: