OpenAI Is Cutting Off Cursor’s Access to Its Models

OpenAI announced on Friday, August 28, 2026 that it intends to wind down the contract that supplies its models to Cursor, the AI code editor acquired by SpaceX earlier this month. The proposed shutoff date is November 12, 2026 — and “proposed” is OpenAI’s own word, repeated throughout its announcement. As of publication, nothing has been cut off: GPT-5.6 Luna, Sol, and Terra are all still listed in Cursor’s own model documentation.

If your team writes code in Cursor, the practical question is not the corporate drama — it is what actually stops working if the proposed date holds, and what does not. The short version: three of the eleven models currently listed in Cursor’s own pricing docs are OpenAI’s, and OpenAI’s published workarounds — bring your own API key, the Codex IDE extension, or an AI gateway — cover Cursor’s local Chat and Agent features only. Tab autocomplete, Auto routing, Cloud and Background Agents, Automations, the Cursor CLI, and the Cursor API/SDK are all explicitly outside that safety net.

The August 28 event itself is documented here from OpenAI’s own first-party material — its announcement and the Help Center article it published as part of the same rollout — cross-checked against Cursor’s evergreen documentation. It covers what was announced, the contract mechanism behind it, the per-surface matrix of what bring-your-own-key does and does not restore, and what a Cursor team should do in the roughly 76-day proposed window.

The announcement came in a post on OpenAI’s own site titled “Our decision on Cursor following its acquisition by SpaceX,” published late Friday, August 28, 2026. Stripped to its operational core, it says four things. OpenAI intends to wind down the contract that provides its models to Cursor. The proposed shutoff date is November 12, 2026. OpenAI is “giving the maximum notice provided by our contract” so developers keep access as long as possible. And OpenAI will not provide its future models to Cursor in the meantime — in its words, “we’ve decided to hold the contract cancellation to the latest date we can while not providing future models to Cursor.”

The context is the acquisition that closed two weeks earlier: SpaceX completed its roughly $60 billion all-stock acquisition of Cursor’s parent Anysphere on August 14, 2026 — we covered the deal itself and the close and Cursor’s first post-close product move when they happened, so this post will not re-narrate the deal.

Two things are worth reading precisely. First, the tone toward Cursor itself is warm: OpenAI says it has “worked with Cursor for nearly four years” and has “enormous respect for their team, their product, and what they’ve built for the developer community.” The stated problem is the new parent company, not the product. Second, every reference to November 12 carries the qualifier “proposed.” The companion Help Center article goes further: “Cursor may choose to end access sooner. We will share the official termination date as soon as it is confirmed between both companies.” A team planning around a hard, agreed date is planning around a date that does not exist yet — in either direction.

A change-of-control clause is a standard contract term that lets one party revisit or cancel an agreement when the other party’s ownership changes — the logic being that you signed a deal with one company and now face a different one. That is the specific mechanism OpenAI says it is invoking here. In its own words, its “custom agreement with Cursor gives us a limited time window to cancel it after a change of control.” SpaceX’s acquisition of Anysphere closing on August 14 opened that window; the August 28 announcement is OpenAI stepping through it.

The contract in question is a bespoke agreement between two large counterparties, and OpenAI is the only party that has described the clause publicly. That matters for anyone trying to generalize from this event: the precedent this sets is narrower than it first appears — it tells you what can happen inside large custom model-supply deals.

The timing arithmetic is simple and worth stating: August 28 to November 12 is roughly 76 days. OpenAI frames that as “the maximum notice provided by our contract” — the most runway the agreement lets it give. That framing cuts both ways. It is genuinely long notice by the standards of model-supply disputes, as the Windsurf comparison in section 07 shows. It is also a reminder that the ceiling on the notice was written into a contract neither you nor Cursor’s users have ever seen.

OpenAI’s stated rationale is a single sentence, and it names names:

The post backs that sentence with two specific exhibits, both linked from OpenAI’s own text. First, after Musk acquired Twitter, the company broke the terms of OpenAI’s contract, per New York Times reporting from April 2023 that OpenAI cites directly. Second, OpenAI points to Forbes coverage from April 2026 of Musk admitting under oath that xAI — now also part of SpaceX — had violated OpenAI’s terms of service. Both citations predate the announcement by months or years; OpenAI chose its evidence from the existing record rather than alleging anything new about Cursor itself.

The ownership chain that makes those exhibits relevant to a code editor: X Corp merged into xAI in March 2025, and SpaceX acquired xAI in February 2026. When SpaceX then bought Anysphere in August, Cursor acquired the same ultimate parent as the two entities OpenAI accuses of prior contract violations. That lineage — not anything Cursor did — is the whole of OpenAI’s stated case.

Buried in the announcement is a forward-looking justification that most coverage skipped. OpenAI ties the decision partly to “a new level of accountability to ensure our upcoming model, Astra, is being used in accordance with our terms.” Astra is OpenAI’s unreleased frontier model — a codename unrelated to Google DeepMind’s similarly named Project Astra — and OpenAI has been publicly cautious about its capabilities: its own post on frontier cyber capabilities discusses preparing for models whose cyber capability level it cannot yet rule out as critical.

Read carefully, the Astra framing is a general accountability bar for large custom partnerships, not a Cursor-specific accusation — OpenAI frames it as a standard it needs to be able to enforce across the board, and the causal language deserves to stay soft. But it is the most interesting sentence in the post for anyone forecasting model supply, because it implies the bar for who gets privileged access to OpenAI’s next frontier model is rising generally. On that reading, the Cursor decision is one visible output of a broader tightening: as claimed capabilities go up, the willingness to route those capabilities through partners whose compliance OpenAI cannot verify goes down. Teams building on any large custom model-supply arrangement — not just Cursor’s — should note the direction of travel.

Alongside the decision post, OpenAI published a Help Center article as part of the same rollout, laying out three paths for developers who want to keep using OpenAI models inside Cursor after the proposed cutoff. This is where the most operationally important fact of the whole story lives — and it is the one almost no coverage walks through. Bring-your-own-key sounds like a full replacement. Per OpenAI’s own article, it is not: “Cursor supports using your own OpenAI API key for OpenAI models in its local Chat and Agent features. Requests are sent through your OpenAI API account and billed separately from your Cursor or ChatGPT subscription.”

The exclusion list that follows is verbatim and exhaustive: “Your OpenAI API key applies only to supported local Chat and Agent requests. It does not apply to: Cursor Tab and autocomplete; Auto model selection and routing; Cloud or Background Agents; Automations; Cursor CLI; Cursor API or SDK; Cursor-provided or proprietary models.” Gateway credentials — routing through Amazon Bedrock, Azure, or an OpenAI-compatible gateway — carry exactly the same scope, per the same article. The matrix below combines OpenAI’s per-surface list with the confirming detail from Cursor’s own evergreen bring-your-own-key documentation, which predates the announcement.

Sources for the table: OpenAI Help Center, “Using OpenAI models in Cursor,” published as part of the same rollout as the August 28 decision post; Cursor Docs, “Bring your own API key.” If your team uses OpenAI models through any surface in the lower group, neither company documents an in-Cursor workaround as of this writing.

Two billing traps hide inside the covered rows. First, “A ChatGPT subscription does not include OpenAI API usage. If you use an API key, usage is billed through your OpenAI API account” — OpenAI’s words. A team on ChatGPT Plus or Pro still needs a separate, metered OpenAI API account with billing configured before a BYOK key does anything. Second, on Cursor’s Teams and Enterprise tiers, Cursor’s own pricing docs list a Cursor Token Rate of $0.25 per million tokens that applies to included, on-demand, and BYOK usage alike — so a BYOK request can carry a small Cursor-side surcharge on top of your OpenAI API bill.

And one compliance trap: Cursor’s BYOK page states that its Zero Data Retention guarantee does not travel with your key — “Your data handling follows the privacy policy of your chosen provider (OpenAI, Anthropic, Google, Azure, or AWS). If your team relies on Zero Data Retention, use Cursor’s built-in models instead.” A team that adopted Cursor partly for its ZDR posture loses that guarantee for OpenAI traffic the moment it switches to BYOK.

The scale of the disruption is easy to overstate, and Cursor’s own models and pricing documentation is the corrective. As retrieved after the announcement — unchanged, since the proposed cutoff has not occurred — Cursor lists eleven models across two pools. The “Cursor Models” pool, with the larger included-usage allowance, holds Grok 4.6, Grok 4.5, and Cursor’s own Composer 2.5. The “Other Models” pool, billed at API rates, holds three Anthropic models, two Google models, and the three OpenAI models this announcement puts at risk.

Source: Cursor Docs, “Models & Pricing,” as listed after the August 28 announcement. Eleven listed models, three affected — every model in the Cursor Models pool and every non-OpenAI line in the Other Models pool is untouched by this specific announcement, since OpenAI is the only provider being wound down. Cursor’s pricing page describes its hosting generically — “Models are hosted by the model provider, a trusted partner, or Cursor” — and does not reference the wind-down anywhere in its model docs. Notably, no Cursor-owned blog post or changelog acknowledging the change could be located as of this writing, so the roster above reflects the pre-cutoff status quo.

One angle worth interpreting rather than just counting: the affected pool is the expensive one. All three GPT-5.6 variants sit in the pool billed at API rates, while Cursor’s included-usage economics are built around Grok and Composer — models supplied by Cursor’s own new corporate family. If the proposed cutoff holds, Cursor’s model menu tilts further toward vertically integrated supply, and the question of which model your AI tool is actually running gets more important for every team that assumed the picker would always be provider-neutral.

The reflexive comparison is June 2025, when Anthropic limited Windsurf’s direct access to Claude models — TechCrunch’s own headline verb was “limiting,” not “banned” — cutting off first-party access to Claude 3.5 Sonnet and Claude 3.7 Sonnet with less than five days’ notice, while Windsurf never had first-party access to Claude 4 at all. Windsurf CEO Varun Mohan told TechCrunch at the time: “We have been very clear to Anthropic that this is not our desire — we wanted to pay them for the full capacity.” Anthropic framed it as compute allocation — “We’re prioritizing capacity for sustainable partnerships that allow us to effectively serve the broader developer community,” a spokesperson said — against the backdrop of reported OpenAI acquisition talks for Windsurf, which led Anthropic co-founder Jared Kaplan to remark that “I think it would be odd for us to be selling Claude to OpenAI.”

The two events rhyme, but they are structurally different, and flattening them into “this keeps happening” obscures what each one teaches. The comparison in full:

What survives the comparison is the underlying dynamic, and it is the durable lesson of both events: a coding tool’s model supply is a lever its infrastructure providers can pull when ownership or competitive alignment shifts. Windsurf’s supply moved when its rumored acquirer was Anthropic’s chief rival; Cursor’s is proposed to move because its actual acquirer is a company OpenAI says it cannot trust with its terms. Different mechanisms, same exposure. We track every dated instance of this pattern in our running ledger of model-supply cutoffs, which this event now joins.

The window is proposed, which means it can shrink — OpenAI’s own guidance says Cursor may choose to end access sooner. The sensible posture is to treat November 12 as the latest plausible date, not the guaranteed one, and to sequence the work accordingly.

First, inventory your actual OpenAI usage in Cursor. The surface matters more than the model. A developer who occasionally picks GPT-5.6 Sol in local Chat can switch to BYOK in an afternoon. A team whose Automations, Background Agents, or CLI scripts are pinned to a GPT-5.6 model ID has no documented in-Cursor workaround at all and needs to either re-pin those workflows to an unaffected model or move them out of Cursor entirely. If you have production automations pinned to specific model IDs anywhere, this is exactly the scenario a one-page model-pinning policy exists to catch.

Second, decide the fallback per surface, not globally. For local Chat and Agent, the choice is BYOK (with a funded OpenAI API account — a ChatGPT subscription does not count), the Codex extension as a parallel agent, or simply selecting an unaffected model. For Auto routing, note that Cursor’s own Auto router sits on the not-covered list — Auto-routed requests cannot fall back to your OpenAI key, so if Auto has been quietly serving you OpenAI models, its behavior may change without any action on your part. For compliance-sensitive teams, re-read the Zero Data Retention caveat before reaching for BYOK as the default answer.

Third, put a re-verification date in the calendar. The official termination date is explicitly to be confirmed between the two companies. Model sunsets with moving dates are a recurring operational hazard — our model retirement calendar exists for exactly this kind of sunset, and this one deserves an entry with a weekly check rather than a single November reminder. Teams that want help pressure-testing their model-supply exposure — mapping which workflows depend on which provider, and what a provider-neutral fallback actually costs — do this kind of audit with our AI transformation practice as a scoped engagement rather than a scramble.

Looking forward: whether or not November 12 holds, the direction is set. OpenAI has now shown it will use contract mechanics to withhold both current and future models from a distribution partner whose ownership it distrusts, and it has tied that posture to the accountability bar for its next frontier model. The rational response for any team building on AI coding tools is not to pick a safer vendor — there may not be one — but to keep per-surface model portability cheap: know what you run, pin deliberately, and rehearse the swap before a provider makes you perform it live.