On August 28, 2026, OpenAI notified SpaceX of its intent to terminate the agreement under which it supplies models to the AI coding environment Cursor. The proposed cutoff date is November 12, with a transition period for current models, but future models will not be made available to Cursor. The trigger was Cursor's acquisition by SpaceX on August 14. Cursor's value proposition as a development environment where users can freely choose models now depends not just on individual model performance, but on the contractual relationships between the companies supplying those models.
The impact is not a shutdown of the Cursor product itself. Grok, Composer, Claude, and Gemini will remain available. The real question is how many of the features that have relied on OpenAI models—accessed through Cursor's contracts and routing—can be migrated to alternative channels.
November 12 is not a fixed deadline
November 12, 2026 is not a confirmed end date but rather a transition deadline proposed by OpenAI; Cursor's access could end even sooner. OpenAI says it will provide further notice once both companies have formally agreed on an end date. In other words, there is no guarantee that the same model lineup will remain in place until November 12.
That said, the treatment of the transition period differs from that of new models. Under OpenAI's proposal, access to the OpenAI models Cursor currently uses will continue through the transition period. Future models, however, will not be provided to Cursor at all. This does not mean that Astra, currently in development, will ever become available through Cursor.
The two companies' partnership spans roughly four years. Cursor co-founder Michael Truell said on August 29 that OpenAI models account for about 5% of Cursor's user traffic, and that the company is in discussions with OpenAI toward a resolution. However, he did not disclose the measurement period or denominator behind that 5% figure. It cannot be translated into a ratio of token volume, revenue, or customer count.
API keys can't fill every gap
As a migration strategy, OpenAI has pointed to three paths: users' own OpenAI API keys, the Codex IDE extension, and compatible gateways. None of these fully preserves all of Cursor's current functionality in its present form. This is a provisional breakdown as of publication. OpenAI itself acknowledges that Cursor's features and availability may change, so users should check Cursor's latest official documentation when migrating.
| Migration path | What continues | What is not replaced |
|---|---|---|
| Personal OpenAI API key | Corresponding local Chat and Agent in Cursor desktop | Tab and autocomplete, Auto, cloud execution, Automations, CLI, API/SDK |
| Codex IDE extension | Codex's agentic experience running in a separate panel within Cursor | Underlying models for Cursor Chat, Agent, Tab, Auto, and Cloud Agents |
| OpenAI-compatible gateway | Corresponding local Chat and Agent | Tab, Auto, Cloud/Background Agents, Automations, CLI, API/SDK |
Continued use of OpenAI models via a personal API key is limited to local Chat and Agent in Cursor desktop; Tab, Auto, Cloud/Background Agents, Automations, CLI, API, and SDK are not covered. These features rely on models that Cursor itself supplies or routes. In other words, while in-editor conversations can be preserved, simply pasting in the same key does not restore cloud agents or automatic model selection.
Billing also changes. The OpenAI API is billed separately from Cursor or ChatGPT subscriptions, and the available models and limits follow the settings of the user's API organization. With compatible gateways, users must also confirm not only that the gateway supports the API format Cursor expects, but also whether model-specific controls are passed through. Rather than simply matching model names, enterprises need to test whether they can bring authentication, logging, and billing back under unified governance.
Why OpenAI didn't wait to establish a breach by Cursor
OpenAI invoked a change-of-control termination clause contained in its individual contract with Cursor. Because the window for exercising this right is limited, OpenAI says it will keep current models available through the latest date permitted under the contract, while stating it cannot be confident that SpaceX will use its technology within the bounds of its terms. This is not an announcement declaring that Cursor itself has committed a violation. Rather, it reflects a judgment that, given the new ownership following the acquisition, OpenAI is unwilling to accept the future supply risk involved.
OpenAI's publicly available business terms prohibit model extraction and circumventing safety measures, and also restrict using outputs to develop AI models that compete with OpenAI. However, the actual text of OpenAI's individual contract with Cursor has not been made public. It is not possible to conclude from the general terms alone that Cursor has violated anything.
OpenAI's decision to withhold future models is also tied to its assessment of Astra's capabilities. On August 7, the company announced that Astra's agentic coding and cybersecurity capabilities had advanced significantly, to the point that it could not rule out reaching the "Critical" threshold under its Preparedness Framework. This is a preliminary internal assessment, not a confirmed finding that Astra has reached Critical. By comparison, the previous GPT-5.6 Sol was rated "High."
The bar for Critical is strict: it refers to a level at which a model can discover and develop zero-day exploits that function across numerous robust, real-world systems without human assistance, or can devise and fully execute novel attack strategies from only high-level goals. As its models' capabilities rise, OpenAI is also demanding tighter controls over the channels through which it supplies third parties. Withholding Astra from Cursor is a concrete contractual example of that control policy in action.
Cursor has become part of SpaceX's vertical-integration push
The relationship between SpaceX and Cursor did not begin with the August acquisition. According to SpaceX's prospectus, the two companies signed a compute agreement on April 19, 2026. Under it, SpaceX provides Cursor with GPU clusters used to develop, train, and improve AI models. In exchange, Cursor contributes personnel, data, and technical expertise, bringing development workflows, prompts, specifications, and code into the collaboration.
The prospectus also notes that the arrangement is intended to improve SpaceX's existing models, including Grok, and that the companies may jointly develop models in the future—with both parties holding broad usage rights to any such jointly developed models. SpaceX has treated Cursor as part of a broader strategy to vertically integrate everything from compute to models to applications.
That said, the scope of "data and datasets" referenced in the public filing is not specified. There is no statement indicating that Cursor users' inputs, companies' proprietary code, or OpenAI model outputs feed into training Grok. While OpenAI's concerns and the structure of the SpaceX-Cursor partnership are both verifiable, inferring the actual data pathways involved would go beyond the available evidence.
Even at 5%, the "neutral infrastructure" problem doesn't go away
Cursor's current pricing documentation groups Grok 4.6, Grok 4.5, and Composer 2.5 under "Cursor Models," while placing Claude, Gemini, and the GPT-5.6 family under "Other Models." In Auto mode for Teams and Enterprise plans, the Cursor Router selects a model for each request. Cursor functions both as a platform that lists models side by side and as the operational infrastructure that decides which model runs under which conditions.
Other model suppliers have responded differently. Anthropic co-founder Tom Brown called Cursor a longtime trusted partner and said Anthropic would increase the compute supporting Claude. This isn't a situation where third-party models disappear en masse simply because OpenAI is stepping back—Cursor's range of options should remain intact for the time being.
Still, the weight behind Truell's description of Cursor as "neutral infrastructure" matters. Even if usage of a particular model sits at just 5%, supply agreements can still reshape the design of Auto, cloud agents, and the CLI. What enterprises need to verify isn't simply whether OpenAI models can be called from some separate screen. It's whether the same work can be shifted to a different channel while preserving existing approval workflows, audit logs, and cost controls.
Whether Cursor can continue to function as a platform offering a genuine choice among multiple models can't be determined just by looking at the model roster on the eventual, formal end date. The real answer will come from whether Auto and cloud features let users choose competing models under the same conditions—and from whatever migration measures are presented before access is finally cut off.