Signals Inbox·August 30, 2026·AI DevTools

Why is OpenAI cutting Cursor access?

OpenAI is cutting Cursor’s wholesale access because SpaceX ownership changed the trust equation around a major model-distribution deal, just as OpenAI’s models are becoming more sensitive and Codex gives it a direct route to developers.

Q1What is OpenAI actually cutting off in Cursor?

OpenAI is currently winding down Cursor’s company-level access to OpenAI models, while individual developers still have several ways to use OpenAI inside the Cursor editor.

OpenAI has proposed ending the custom contract under which Cursor buys and distributes OpenAI models to its users. During the transition, Cursor can keep serving the OpenAI models already covered by that agreement, although future OpenAI models will no longer be added through the same relationship.

For a developer, the situation is less dramatic than “GPT is disappearing from Cursor.” OpenAI’s latest guidance says users can bring their own OpenAI API key into supported local Cursor Chat and Agent workflows. They can also run the Codex IDE extension directly inside Cursor with a ChatGPT subscription or OpenAI API account. Compatible gateways provide another route.

Those workarounds do not reproduce everything Cursor currently does. Cursor’s own documentation says personal API keys cover supported chat models, while features such as Tab still rely on Cursor’s infrastructure. Cloud Agents also use a model set controlled by Cursor.

The real cut therefore happens at the platform level. OpenAI is removing Cursor’s ability to package OpenAI models broadly as part of Cursor’s own service, while continuing to sell OpenAI access directly to developers.

How OpenAI access changes inside Cursor

Q2Why did OpenAI decide to cut Cursor access now?

SpaceX’s takeover of Cursor gave OpenAI a narrow contractual opportunity to leave, making the ownership change the clearest reason the decision happened now.

OpenAI says its custom Cursor agreement contains a change-of-control clause with a limited cancellation window. Cursor officially became part of SpaceX roughly two weeks before OpenAI announced that it wanted to wind down the contract.

That timing is important because OpenAI’s concerns about Elon Musk’s companies existed earlier. OpenAI had already clashed with Twitter over an old data agreement, and Musk had already acknowledged under oath that xAI partly used OpenAI technology while developing its own models. Neither event immediately ended Cursor’s access.

The ownership change forced a decision. OpenAI could continue treating Cursor as a long-standing independent customer, or reconsider the agreement now that Cursor belonged to a group building rival frontier models.

OpenAI chose the second option.

Cursor cofounder Michael Truell says the companies are still talking, so the final operating details could change. But the reason the issue surfaced right now is unusually clear: the contract gave OpenAI a short window in which new ownership could trigger an exit.

Q3Did Cursor actually break OpenAI’s rules?

Cursor itself has not been accused of breaking OpenAI’s rules, and OpenAI’s public explanation focuses instead on what it expects from Cursor’s new corporate parent.

That distinction gets lost easily.

OpenAI says it worked with Cursor for nearly four years and praised both the team and the product when announcing the planned termination. There is no public allegation that Cursor secretly trained models on prohibited OpenAI outputs, violated usage restrictions or broke its existing contract before the ownership change.

OpenAI’s argument comes from previous dealings with other companies controlled by Musk.

One example involves Twitter. OpenAI says that after Musk bought the company, Twitter broke the terms of a data agreement between the two businesses.

The more relevant example involves xAI. During court testimony earlier this year, Musk said xAI had “partly” used OpenAI technology while training its own AI models. OpenAI’s business terms prohibit customers from using outputs to develop AI models that compete with OpenAI, outside a few narrow exceptions.

That gives OpenAI a concrete reason to worry about future compliance, but it does not establish wrongdoing by Cursor.

For Cursor’s reputation, that is an important difference. OpenAI is reacting to who now controls the company and what sits around it, rather than presenting evidence that Cursor itself cheated.

Q4Why does xAI’s OpenAI distillation matter for Cursor?

xAI’s earlier use of OpenAI technology makes Cursor’s new position much harder for OpenAI to treat like a normal customer relationship.

Cursor is increasingly involved in model development itself.

Earlier this year, Cursor said its Composer models had become limited by the amount of compute available for training. It then partnered with SpaceXAI to use the Colossus infrastructure for much larger training runs. Cursor later released Grok 4.6 together with SpaceXAI, describing it as a first-party Cursor model built for coding and broader knowledge work.

OpenAI now has to think about Cursor in that context. The same corporate group can distribute AI coding products, train frontier models and compete directly with OpenAI.

The legal boundary is also unusually relevant. OpenAI’s current Services Agreement says customers cannot use OpenAI output to develop competing AI models, except for specific permitted cases. Musk’s testimony gives OpenAI evidence that another part of the same group has crossed that boundary before.

We have found no evidence that Cursor used OpenAI output to train Composer or Grok 4.6. That would be a much stronger accusation and the available evidence does not support it.

The problem for OpenAI is simpler: trusting the separation between product usage and rival-model development now requires more confidence than OpenAI appears willing to give.

Q5Is OpenAI really worried about Astra being used through Cursor?

OpenAI’s Astra concerns look substantive because the company had already tightened security around the model before the Cursor dispute appeared.

OpenAI recently disclosed that internal Astra evaluations showed enough progress in agentic coding and cybersecurity that it could no longer rule out “Critical” cyber capability under its Preparedness Framework. Previous models including GPT-5.6 Sol had remained at the lower “High” threshold.

That is a big jump in practical risk.

Under OpenAI’s own definition, the Critical threshold includes capabilities such as finding and developing functional zero-day exploits against hardened real-world systems or executing sophisticated attacks with little human help.

OpenAI responded internally with tighter network access, stronger monitoring, more isolated testing environments and restrictions on Astra work that did not meet the new controls.

Then Astra appeared directly in OpenAI’s explanation for the Cursor decision. OpenAI said more capable models create a higher level of responsibility around ensuring that customers follow its terms.

So Astra should not be treated as a random safety excuse added to a commercial fight. OpenAI had already documented the model’s unusual cyber capabilities and tightened its own internal controls.

The SpaceX relationship made those concerns harder for OpenAI to ignore precisely when the model itself became more sensitive.

Q6Is the Elon Musk and Sam Altman feud really driving the Cursor cutoff?

The Musk-Altman feud probably makes compromise harder, but the evidence gives OpenAI stronger reasons than personal hostility for cutting Cursor access.

The relationship between Musk and OpenAI is obviously poisonous. Years of lawsuits, public accusations and competing narratives about OpenAI’s founding have removed much of the trust that normally helps large commercial partnerships survive ambiguous situations.

Still, the Cursor decision follows a specific contractual path rather than a random escalation in the feud.

OpenAI had continued supplying Cursor while Musk was already fighting the company in court. It also tolerated a world in which Cursor offered competing models alongside GPT. The direct relationship only came under threat when OpenAI had to decide whether to keep the same custom agreement under a different owner.

That makes the feud more useful as context than as the main explanation.

Personal history probably pushes OpenAI toward the least trusting interpretation whenever a contract leaves room for judgment. But even if Musk and Altman liked each other, a frontier-model supplier would still have to think carefully before giving a rival model group broad wholesale access to increasingly sensitive technology.

Q7Is OpenAI also trying to push Cursor users toward Codex?

OpenAI clearly gains from making Codex the easiest way to use OpenAI models for coding, and that commercial upside is now too large to ignore.

The developer market has moved quickly.

A recent JetBrains survey covering more than 15,000 professional developers found that Codex usage at work rose from 3% in January to 16% in its May-to-July survey period. Cursor moved from 18% to 12% over the same comparison, while Claude Code reached 39%.

That does not prove Codex stole those Cursor users. Developers often use several tools at once, and Claude Code grew much more than either product.

But OpenAI now owns a coding product with serious distribution of its own. JetBrains also recently made Codex its recommended starting agent after benchmarking solve rate, cost and latency on hundreds of real software-engineering tasks and then validating the choice with an A/B test.

OpenAI’s latest Cursor guidance fits that strategy perfectly. One of the recommended ways to keep using OpenAI inside Cursor is simply to install Codex.

The editor can remain Cursor while the AI relationship moves to OpenAI.

That is commercially attractive for OpenAI because authentication, usage and the agent experience increasingly belong to Codex rather than Cursor. Capacity pressure may provide a secondary benefit, but it does not explain the pattern nearly as well: OpenAI is still inviting developers to consume the same models through its own channels.

AI coding tool usage in the recent JetBrains survey

Q8How dependent is Cursor on OpenAI models right now?

Cursor currently looks surprisingly independent from OpenAI: Michael Truell says OpenAI models account for only about 5% of Cursor user traffic.

That number sharply lowers the risk of an immediate collapse.

If the figure reflects actual model requests, roughly 95% of Cursor traffic already runs through other models. Cursor has Anthropic and Google models available, plus its own Composer family and the Grok models it now develops with SpaceXAI.

The 5% figure still needs to be interpreted carefully. Traffic share tells us nothing about the economic value of the users generating those requests. OpenAI-heavy customers could be unusually large enterprises, and some developers may deliberately save a model such as GPT-5.6 Sol for difficult tasks while using cheaper models for routine work.

It also does not measure the value of optionality. Part of Cursor’s appeal has always been that users can switch among frontier providers without changing their coding environment.

Even with those caveats, 5% is too small to describe Cursor as structurally dependent on OpenAI.

Cursor is losing an important model supplier. It is nowhere close to losing the model stack on which the whole product runs.

Q9Can Cursor replace OpenAI with its own AI models?

Cursor can replace much of the practical OpenAI usage today because its first-party models have moved far beyond the experimental stage.

The progression has been fast. Composer 1.5 increased reinforcement-learning compute by more than twentyfold compared with the original model. Composer 2 then added continued pretraining and scored 73.7 on SWE-bench Multilingual and 61.7 on Terminal-Bench in Cursor’s published evaluations. Composer 2.5 pushed the training process further with harder reinforcement-learning environments and synthetic data.

Grok 4.6 takes the strategy another step.

Cursor and SpaceXAI say Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index. Cursor’s published benchmark set also shows Grok 4.6 ahead on some agentic coding and knowledge-work evaluations while GPT-5.6 Sol remains stronger on others.

Those numbers come partly from Cursor’s own comparison and should be treated accordingly. They show that Cursor has reached roughly the right performance neighborhood, rather than proving Grok is universally better than GPT.

The economics favor first-party models too. Cursor’s pricing now separates a generous “Cursor Models” pool containing Grok and Composer from a third-party model pool containing providers such as OpenAI, Anthropic and Google.

So replacing some GPT traffic does more than reduce supplier risk. It can also give Cursor more control over cost, routing and product behavior.

Q10Could losing OpenAI still hurt Cursor’s enterprise business?

Losing direct OpenAI access can still create real enterprise friction even if OpenAI represents only a small share of Cursor’s overall traffic.

Cursor is now a very large business product. Forbes reported that its annualized revenue climbed from about $2 billion in February to $3 billion in late April and more than $4 billion by early June. Cursor told Forbes that roughly three-quarters of that run rate came from business customers.

At a $4 billion run rate, a 75% business mix implies around $3 billion of annualized revenue tied to companies.

Those customers care about more than which model handles the median request. Large teams can have internal benchmarks built around specific OpenAI models, security approval for certain vendors, negotiated API agreements or workflows that depend on switching models for different tasks.

OpenAI’s personal-key workaround only covers part of Cursor. Cursor’s documentation says BYOK works for supported chat models, while first-party infrastructure remains necessary for features such as Tab. Cloud Agents also run through a model selection controlled by Cursor.

For an individual developer, replacing GPT may be a small inconvenience. For a company with thousands of engineers and centrally managed workflows, changing model routing can involve procurement, security and tooling work.

That is where the OpenAI cutoff can hurt Cursor more than the 5% traffic figure initially suggests.

Cursor business metrics

Q11Will Anthropic cut Claude access to Cursor next?

Anthropic currently appears to be moving in the opposite direction: its chief compute officer says the company plans to increase compute supporting Claude models inside Cursor.

That is one of the most important developments since OpenAI announced its decision.

Anthropic cofounder Tom Brown publicly described Cursor as a trusted partner going back to Sonnet 3.5 and said Anthropic would continue increasing compute for Claude in Cursor. Other people working on Claude Code have publicly echoed support for continuing the relationship.

For Cursor, this changes the practical risk considerably.

Claude Code is currently the most widely adopted specialist AI coding tool in JetBrains’ latest survey, and Claude models remain central to many professional coding workflows. Losing OpenAI and Anthropic together would create a serious supplier problem. Losing OpenAI while Anthropic actively expands support is much easier to absorb.

The contrast also tells us something about OpenAI’s justification.

Anthropic has its own reasons to worry about model extraction and competitive use. Yet it currently considers the Cursor relationship worth maintaining even under SpaceX ownership. That means OpenAI’s decision cannot be treated as an inevitable industry-wide response to Cursor’s new corporate structure.

OpenAI is making a stricter trust judgment than Anthropic is, at least for now.

Q12Are AI model companies starting to cut off coding tools they compete with?

Frontier AI companies have already shown that model access can disappear when a coding platform becomes too closely tied to a major competitor.

The clearest earlier case involved Windsurf.

When OpenAI was reportedly preparing to acquire Windsurf in 2025, Anthropic reduced Windsurf’s direct access to Claude models. Anthropic cofounder Jared Kaplan explained the logic very plainly at the time: selling Claude to a company about to belong to OpenAI would be an odd long-term relationship.

Windsurf had to rely more heavily on alternative ways of delivering Claude.

That episode looks much more relevant today than it did a year ago. Cursor has now landed in a similar position from the opposite direction: a formerly independent coding product becomes attached to a frontier-model competitor, and one major supplier rethinks direct access.

There is an important difference. Anthropic is currently sticking with Cursor despite SpaceX ownership, which shows these decisions are neither automatic nor universal.

Still, the direction is getting hard to miss. AI coding products started as relatively neutral layers sitting above several model companies. As the largest coding products build their own models and the largest model companies build their own coding agents, that neutral middle ground is getting smaller.

Model availability is increasingly becoming a competitive weapon as well as a technical choice.

Q13Could OpenAI’s Cursor cutoff backfire?

OpenAI risks encouraging Cursor and other AI coding companies to become even less dependent on outside frontier-model providers.

Cursor already has the ingredients to respond that way.

It has trained several generations of Composer, gained access to far more compute through SpaceXAI and released Grok 4.6 as a first-party Cursor model. Its pricing also gives users more included usage for Cursor models than for third-party ones.

A supplier cutoff gives Cursor one more incentive to improve that stack and route more traffic toward models it controls.

The wider lesson could be even more important. Independent AI applications became valuable partly because they let users choose the best model without changing products. If access to a frontier model can disappear after an acquisition, partnership or strategic disagreement, companies building on those APIs have a reason to diversify earlier.

We have already seen the same basic vulnerability with Windsurf.

OpenAI reduces one risk by keeping direct developer routes open. A Cursor user who really wants GPT can still use compatible OpenAI access or run Codex inside the editor. That prevents the decision from becoming a total model blockade.

Even so, OpenAI is trading distribution for control. Today that looks manageable because Codex has become strong enough to reach developers directly. Over time, the cost could be a developer ecosystem that trusts model suppliers less as neutral infrastructure.

Q14So why is OpenAI cutting Cursor access?

OpenAI is cutting Cursor’s direct model access mainly because it no longer trusts the corporate structure around Cursor enough to keep supplying frontier models through a large custom contract, while the rise of Codex makes that decision commercially much easier.

The evidence supports that explanation far more than a simple personal feud or a shortage of compute.

OpenAI had worked with Cursor for years. The relationship became a problem once the contractual change-of-control window opened and OpenAI had to judge Cursor as part of a group that also develops rival frontier models. OpenAI could then point to previous contract trouble with Musk-controlled companies and, more seriously, xAI’s acknowledged use of OpenAI technology during model development.

As seen above, OpenAI is also tightening access around Astra at a time when its own evaluations suggest much stronger cyber capabilities. That raises the cost of being relaxed about who receives new models through large third-party integrations.

The competitive incentive strengthens the decision further. Codex usage has risen quickly, JetBrains now recommends Codex as its default starting agent, and OpenAI can keep serving developers directly without depending on Cursor for distribution.

Cursor should survive the move comfortably based on what we know today. OpenAI represents about 5% of its traffic, Cursor has increasingly capable first-party models, and Anthropic has just said it plans to expand rather than reduce Claude compute for Cursor.

So the strongest answer is that OpenAI is drawing a new trust boundary around wholesale access to its frontier models. Cursor users remain valuable OpenAI customers. SpaceX-controlled Cursor itself is a partner OpenAI currently no longer wants to trust with the same level of access.