OpenAI Claims 10,000 AI Agents Solved Navier–Stokes in Under 90 Hours — Without Peer Review

OpenAI Claims 10,000 AI Agents Solved Navier–Stokes in Under 90 Hours — Without Peer Review

On Tuesday morning, OpenAI published a blog post claiming that an internal AI model has solved the Navier–Stokes existence and smoothness problem, one of the Clay Mathematics Institute's seven Millennium Prize Problems. But the result is unverified, the $1 million prize remains open — and the announcement has triggered a public priority dispute with two mathematicians who published related results shortly beforehand.

What OpenAI Claims

In the blog post, which AIMag knows only through secondary coverage by Gizmodo, OpenAI claims that an unreleased model described as "significantly more capable than GPT-6 Astra" has solved the problem, which concerns the physics of water and other fluids. The result was supposedly discovered by human researchers collaborating with roughly 10,000 AI agents over just under 90 hours. OpenAI writes that the company does not intend to accept the $1 million prize.

The proof itself is described in coverage by 24ai.no as a vortex motion that spirals inward and stretches out, accelerating continuously while the energy remains finite. According to OpenAI, this would establish that singularities can develop in finite time — thereby confirming what in CMI's terminology are called claims "C" and "D" in the official problem statement.

The Navier–Stokes problem is one of seven Millennium Prize Problems announced by the Clay Mathematics Institute in 2000, each with a $1 million prize for the first correct solution.

What Actually Exists — and What Is Missing

It is important to distinguish between a claim and a validated result. OpenAI has reportedly published a 165-page written paper and a formalization in the Lean proof assistant. According to 24ai.no, citing unnamed sources, the Lean formalization was supposedly verified by GPT-6 Astra in 17 hours. These details cannot be confirmed against primary sources; neither the blog post, the paper, nor the Lean repository is available to us.

What is established is this: as of September 2026, the Millennium Prize for Navier–Stokes remains formally open. The Clay Mathematics Institute has neither confirmed nor validated OpenAI's result, and full peer review from the broader mathematics community is still outstanding.

Fields Medalist Timothy Gowers said the result, if it holds, is undoubtedly "a big moment" — while at the same time emphasizing that he had not read the proof when he spoke, according to 24ai.no.

The Priority Dispute

Shortly before OpenAI's announcement, the mathematicians Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) published related results on forced Euler equations. Buckmaster claims that OpenAI accelerated its work after hearing rumors of their breakthrough, and that work in development stored in OpenAI's Codex model may have been accessible to the OpenAI team. This is an opposing-party claim that cannot be clarified from available sources.

OpenAI presents its own timeline: the work began September 1, after the rumors, and the company first contacted Buckmaster and Alpöge on September 6 — after completing its own proof and Lean verification. In the blog post, OpenAI writes that neither the company's researchers nor the AI agents saw the two researchers' work before it was published, but at the same time concedes that the company "cannot rule out that de-identified data derived from their usage of our products helped improve our models."

The dispute took on a personal turn when Bubeck — whose role or position at OpenAI is not specified in available coverage — wrote in a reply on X on Tuesday that he "cannot regard Levent as an independent academic," after learning that Anthropic's AI models had been used in Alpöge's work with Buckmaster. Bubeck later apologized for a question to Buckmaster that was perceived as threatening, calling the phrasing "an extremely poor choice of words" and "the exact opposite of what I was trying to convey," according to Gizmodo. The notion that Alpöge's use of Anthropic's models makes his independence problematic is thus OpenAI's interpretation, not an established fact.

Who came first, and whether training data played a role, cannot be determined from the available evidence. What is certain is that the dispute itself illustrates something new: priority questions in mathematics have been bitter before, but now the quarrel also concerns who has trained which models on whose behavior.

The Principled Question: Knowing Versus Understanding

Terence Tao, regarded as one of the world's foremost living mathematicians, praised Buckmaster and Alpöge's work as "a remarkable achievement." But he also raised a principled concern: that AI-generated solutions could "pollute" problems as a source of further mathematical progress. The point is the difference between knowing that something is true and understanding why it is true. A machine-verified proof that no human understands formally settles the problem, but may not give the field the mathematical foundation of insight that a proof traditionally provides.

Tao took the consequence further in a comment reported by Gizmodo: if humanity's collective efforts to understand the universe — which depend not only on friendly competition but on the free and open exchange of ideas — are undermined by the competition between "tech titans," the total stock of shared human knowledge is doomed to shrink rather than grow.

Javier Gómez-Serrano is more optimistic in the coverage, but agrees with the main point: a proof no one understands has limited value.

Here lies the real question behind the headlines. Lean verification guarantees that a formal derivation contains no errors — provided the formalization actually captures what it claims. It does not guarantee that humans understand the mechanism, can generalize the technique, or can build on it. That is why CMI's evaluation process is not a formality: the Institute must decide whether the result holds as a solution to the official problem, and the mathematics community must assess whether the proof is actually understandable and reproducible.

What Remains

Three things must happen before this matter can be settled. First, full peer review of the 165-page paper — no independent peer review exists. Second, CMI's own validation processes, which typically take time; the prize remains formally open. Third, the question of whether the broader mathematics community can actually read, understand, and verify the proof — the point raised by Tao and Gómez-Serrano.

In addition, the priority dispute hangs in the air. Buckmaster's claim of rumor-driven acceleration and possible access to Codex-stored work, and OpenAI's admission that de-identified usage data "cannot be ruled out" in the model training, point to an accountability question that runs deeper than this single case: What does it mean for research integrity when the same companies build the tools, own the usage data from their competitors' researchers, and announce breakthroughs on their own blogs? The answer is not in today's sources — but the question is now public, and it is being asked ever more frequently.