Why OpenAI Plans to End Its Direct Model Agreement With Cursor

OpenAI’s announcement about Cursor may look routine, but I think the dispute behind it raises larger questions about model access and competition. In this September 2026 video, I examine the company’s stated reasons, responses from Cursor and Anthropic, and what the decision could mean for developers.

Berman interprets the Friday-evening announcement as an attempt to limit attention; that is his inference, not a stated company motive. On August 28, OpenAI said it intended to wind down its direct agreement supplying models to Cursor after SpaceX acquired it, proposing November 12, 2026 as the shutoff date. The announcement did not describe an immediate, universal ban.

OpenAI said it could not be confident that SpaceX would use its technology within its terms, citing its experience with Musk-controlled companies and past contractual issues. It also said the agreement gave it a limited cancellation window after a change of control and that future models would not be provided under that arrangement. These are OpenAI’s stated reasons, not proof that Cursor misused model outputs.

Berman recounts the history of Elon Musk and Sam Altman at OpenAI: the organization’s nonprofit founding, Musk’s departure, the later capped-profit structure and the emergence of xAI as a competitor. He describes their disagreement over the organization’s direction; the parties have offered competing accounts.

Musk sued OpenAI over its direction in 2024. In May 2026, a federal court dismissed claims in a California case after finding they were brought too late; the outcome should not be generalized to every dispute between the parties. Berman argues that access to multiple models can be valuable for developers and adversarial review.

The video is sponsored by Greptile, a code-review product. Berman describes an example in which its tool flagged a reported issue in an open-source project, and says he uses it on his own projects. This is a sponsor presentation and his endorsement, not an independent comparison of code-review tools.

Returning to OpenAI’s statement, Berman discusses its references to past dealings with Musk-controlled companies. OpenAI said Twitter had broken a contract and cited Musk’s testimony concerning xAI’s use of its models. Those statements concern the company’s stated experience; they do not establish that Cursor itself violated its agreement.

Berman explains distillation in broad terms: a developer can use examples or outputs from a larger model to train another model. He argues this can lower the work needed to create a capable model, while noting that providers may restrict such use in their terms.

He refers to Anthropic’s earlier allegations of large-scale extraction by other labs and draws a possible analogy with OpenAI’s concern. Whether a particular system was trained on prohibited outputs, and how capable the result is, requires evidence; the video’s examples do not prove that Cursor or SpaceX did so here.

Berman interprets OpenAI’s decision as concern that access to model outputs could help a competitor train models. This is his explanation of a possible incentive; OpenAI’s announcement states a broader concern about compliance and past dealings, rather than proving a Cursor distillation incident.

Berman discloses existing business relationships with both Cursor and OpenAI and says he likes both companies. He shows Cursor CEO Michael Truell’s response: Truell said OpenAI models accounted for about 5% of Cursor user traffic and that the companies were speaking about a resolution.

Berman says that 5% traffic does not necessarily mean 5% of revenue and wonders what the traffic measure counts. He discusses an OpenAI representative’s argument that token volume and value are different. Neither the exact denominator nor the economic impact is established by the cited statement.

Berman considers SpaceX’s acquisition of Cursor strategically valuable, arguing that Cursor offers distribution and coding experience while xAI has substantial compute. He speculates that model training and access to computing resources were important motivations and predicts that future Grok models will improve; those predictions are not demonstrated outcomes.

Berman claims Cursor has all users’ coding data available to train future models. That is too broad: Cursor’s published privacy documentation says code is not used for training by Cursor or other model providers when Privacy Mode is enabled. Data handling depends on settings, agreements and policy, and this video does not establish training on all user data.

Berman describes direct native access as the affected relationship and says users may have alternatives such as bringing their own API key or using an extension or compatible gateway. Availability and terms depend on the particular product and integration; OpenAI’s August statement sets out the proposed contract termination, not a guarantee that every workaround will remain available.

Berman compares other restrictions on model access, including Anthropic’s earlier limits involving Windsurf and access by competing labs. His claim that Google bought Windsurf oversimplifies events: Google hired Windsurf’s leadership and licensed technology, while Cognition acquired Windsurf’s remaining business. The examples show providers making different access decisions under different circumstances.

Berman contrasts those earlier restrictions with a statement from Anthropic co-founder Tom Brown that Claude models would continue to be supported in Cursor following the SpaceX acquisition. He sees a tension in the decisions, though they concern different relationships and circumstances.

He recalls a difficult public relationship between Musk and Anthropic and discusses their subsequent engagement and compute arrangements. His explanation of the companies’ motivations and the extent of available idle capacity is interpretation, not independently established evidence.

Berman argues that a coding platform dependent on an outside model supplier faces platform risk. He expects Cursor and other companies to invest more in their own models and predicts that leading AI labs will increasingly build their own models, developer tools and platforms.

In Berman’s view, that vertical integration could make it harder to use several providers’ models in one interface. This is a forecast, not a proven market outcome. He closes by mentioning Musk’s hostile public response and a separate interview with an OpenAI representative.