On 28 August, OpenAI published a note titled “Our decision on Cursor following its acquisition by SpaceX”. The substance of it is a single sentence: OpenAI is winding down the contract that supplies its models to Cursor, and under its proposal, Cursor’s direct access to those models ends on 12 November. It was on the front page of Hacker News within hours, where it was read mostly as the next instalment in the Musk and Altman feud.
That is the least interesting thing about it. We covered the SpaceX acquisition of Anysphere in June, when the question was lock-in and portable context. This is the sequel, and a materially different risk. The models inside your development tooling arrive under a commercial agreement between two other companies. You are not a party to it, you cannot see its termination clause, you cannot invoke it, and here it was pulled for reasons that have nothing to do with you.
TL;DR
- OpenAI is ending its agreement to supply models to Cursor after SpaceX completed its acquisition of Anysphere, with a proposed cutoff of 12 November 2026: roughly eleven weeks of notice on a tool many teams standardised on.
- The stated reason is counterparty trust, not technical: OpenAI says it cannot be confident SpaceX will honour its terms of service, citing X’s conduct under the same ownership.
- Cursor CEO Michael Truell says OpenAI models carry about 5% of Cursor traffic and that talks continue. Real exposure, not existential, which is why it is a useful case to study calmly.
- The precedent is not isolated: Anthropic cut off xAI’s Claude access after xAI staff used Claude through Cursor for internal development, and revoked OpenAI’s Claude API access in 2025 on similar grounds.
- The action for engineering leaders is procurement, not panic: know which of your tools resell someone else’s models, ask for the contractual notice period in writing, and keep the harness separable from the model.
What actually happened
SpaceX announced an all-stock acquisition of Anysphere, the company behind Cursor, valued at around $60bn in June, completing it earlier this month. On 28 August OpenAI notified SpaceX that it intends to wind down the contract under which its models are served inside Cursor, proposing 12 November as the shutoff date.
OpenAI’s stated reasoning is that it cannot be confident SpaceX will use its technology within the terms of service, pointing to its experience of X breaching contract terms after Musk acquired that platform. Michael Truell, Cursor co-founder and now a SpaceX executive, responded publicly that the company is disappointed, that OpenAI models account for roughly 5% of Cursor user traffic, that talks are under way, and that Cursor was one of OpenAI’s earliest customers after seed funding from the OpenAI Startup Fund in 2023.
Note what is absent from that exchange: any consideration of the teams with Cursor in their onboarding docs. Not through malice. They are simply not at the table, because they are not on the contract.
This is not the lock-in story, and that matters
Vendor lock-in is a well-understood risk with well-understood mitigations. You worry about it when a supplier makes leaving expensive: proprietary formats, data gravity, bespoke integrations. It is a risk you can price during selection, and the mitigations are yours to implement.
What happened to Cursor is a different shape. Nothing about the product got worse. No price rose. The company did not fail or change its API. A third party, with which the customer has no relationship, decided to stop supplying an input. Call it counterparty revocation: the supplier of your supplier withdrew, unilaterally, for reasons rooted in corporate ownership rather than anything you or your vendor did.
The tell is that no standard software procurement pack surfaces it. We review a lot of these. They ask about uptime, data residency, breach notification, subprocessors, exit assistance and source escrow. Almost none ask the question that mattered here: which of the capabilities you are selling us do you not own, and what notice would you get if that supply ended?
The detail that should worry you most
Buried in the background is a more instructive incident. Anthropic cut off xAI’s access to Claude earlier this year after finding xAI staff were using Claude for internal development, breaching the clause barring use of the models to build competing systems. The route they used was Cursor.
Access was revoked not because of a flaw in the tool, nor because of anything the tool’s vendor did wrong, but because of who was sitting in some of the seats. The model provider looked through the tool at the end users and made a judgement about them. If your tooling reaches models through a reseller, your continuity depends partly on the conduct of that reseller’s other customers, and on how a provider interprets its own acceptable use policy in a given week. That has now happened three times in about twelve months, involving four of the largest companies in the sector.
What eleven weeks actually costs
Cursor’s exposure looks survivable: five per cent of traffic, first-party models of its own, other providers still connected. The number worth calculating is not Cursor’s, it is yours.
If a model you rely on for a specific task disappeared from your tooling in eleven weeks, what is the actual bill? In our experience it is rarely the licence. It is the accumulated fit: prompts and rules files tuned to one model’s behaviour, review habits built around its particular failure modes, internal evaluation baselines, CI steps that assume a given latency and cost profile, and a fortnight of reduced throughput while the team relearns a tool. None of that appears on an invoice, all of it is real, and it is why “we can just switch” is usually optimism rather than a plan.
The teams handling this well separated two things early: the harness, meaning the editor, the agent runner, the context plumbing and the review workflow, and the model behind it. Where that separation is a genuine architectural property rather than a slide, a supply change is a configuration change. Where it is not, it is an unscheduled migration.
What to do about it
- Inventory the resale layer. Mark every AI-capable tool in your stack as first-party, resold (someone else’s models under the vendor’s brand) or bring-your-own-key. The resold column is your counterparty exposure, and it is usually longer than expected once support desks, analytics and design tools are counted.
- Ask the notice question in writing. At renewal, ask each vendor what contractual notice they would receive if an upstream provider terminated, and what they commit to giving you. A vendor with a good answer has one ready. A vendor that has never considered it has told you something useful.
- Keep a second provider warm, not theoretical. Warm means it has run your real evaluations on your real tasks this quarter, with results written down, and a named person knows how to switch the configuration. An account you opened once is not a fallback.
- Own the artefacts that encode your working knowledge. Rules files, prompt libraries, evaluation sets and agent definitions belong in your repositories under your version control, not solely inside a vendor’s workspace. Cheapest item on this list, and the one most often skipped.
- Read the acceptable use policy as an operational document. The competing-systems clause is standard across the major providers and is now demonstrably enforced. If your company builds anything that could be characterised as competing, that clause deserves an actual reading rather than a checkbox.
- Put counterparty risk in the architecture review. If this supply stopped in ninety days, what would we do and who would do it. If the answer requires a meeting to discover, you have found the gap.
The same clause is in your own product
There is an uncomfortable inversion here for anyone shipping software with AI features, which by now is most of our clients. If you resell inference inside your own product, your customers carry exactly the exposure described above and are trusting you to manage it without ever having asked. Your terms probably promise a feature. Your upstream contract probably permits termination on notice measured in weeks. Somebody eventually reconciles those two documents, and it is better that it is you, before a customer’s procurement team does it at renewal.
The reasonable position is not to avoid resold models. Building on frontier models you do not operate is the correct call for the overwhelming majority of teams, and self-hosting carries its own bill. It is to know the dependency exists, to have it written down, and to have decided in advance what happens when it moves. These relationships are commercial and political as much as technical, and they can change in an afternoon over something that had nothing to do with your product.
REPTILEHAUS builds and operates AI-backed systems that keep working when the market underneath them shifts: model-agnostic architecture with a real abstraction rather than a wrapper, evaluation harnesses that make provider switching a measured decision, agent and workflow infrastructure you own outright, and due diligence on the AI dependencies buried in your existing stack. If you are not certain which parts of your toolchain someone else could switch off, get in touch.
📷 Photo by Jordan Harrison on Unsplash
