A Mathematician Accuses OpenAI of Stealing His Proof

OpenAI Navier-Stokes Millennium Prize proof credit dispute mathematician 2026

The Claim

Navier-Stokes, Solved?

88 hours

Time OpenAI says its model took, per its own account to reporters

~10,000

AI agents OpenAI says worked on the problem in parallel

100 pages

Length of OpenAI's proof, not yet reviewed by outside mathematicians

$1M

Clay Mathematics Institute prize OpenAI says it will not claim

Published: September 9, 2026 | Updated: September 9, 2026 | Category: AI | By Mahesh

OpenAI announced on September 8 that an internal model, described as significantly more capable than its newly released GPT-6 Astra, generated a proof that the three-dimensional Navier-Stokes equations can develop a singularity in finite time, one of the seven Millennium Prize Problems the Clay Mathematics Institute selected in 2000 as among the hardest unsolved questions in mathematics. According to Nature's own reporting on the announcement, OpenAI posted a preprint describing the result directly on its website, and OpenAI mathematician Sebastian Bubeck told reporters, "This is, to me, the spectacular culmination of the arc we have seen over the last 12 months."^[1] Martin Bridson, president of the Clay Mathematics Institute, told Nature, "It is certainly an exciting day, as we contemplate the announcement of major advances in the human understanding of mathematics."

By OpenAI's own account, relayed to reporters on a call, the effort took roughly 88 hours and deployed approximately 10,000 coordinating AI agents running in parallel, with the output checked using Lean, software that formally verifies mathematical proofs step by step, according to reporting from Rolling Out.^[2] OpenAI has stated it does not intend to claim the $1 million Clay Mathematics Institute prize, instead framing the result as evidence of how quickly its most advanced systems are improving, according to Axios' reporting on the announcement.^[3]

Why OpenAI Started Working on This Problem At All

The origin of OpenAI's effort is where this story shifts from a straightforward capability announcement into something more contested. According to Nature's reporting, OpenAI researchers said they had been testing their latest AI prototype against all six unsolved Millennium Problems, then decided to focus their resources specifically on Navier-Stokes on September 1, after hearing rumors that two outside mathematicians were close to a breakthrough on the same problem. Axios' reporting names those two mathematicians directly: Tristan Buckmaster, a mathematician at NYU, and Levent Alpöge, who works at Anthropic, OpenAI's chief rival.

What Buckmaster Says Actually Happened, In His Own Words

On September 7, the day before OpenAI's public announcement, Buckmaster released a four-page personal statement alongside three fluid mechanics papers, laying out his own account of the sequence of events. According to Entrepreneur's reporting, Buckmaster says he and Alpöge had already made major progress on the underlying problem before OpenAI's effort began, and that word of their specific approach reached OpenAI before they could publish, after which the company used it to finish its own proof over a single weekend.^[4] BigGo Finance's detailed account of Buckmaster's statement adds further specifics: he learned on a call with OpenAI that the company's internal model had generated a roughly 100-page proof, and that the technical approach OpenAI had taken showed a high degree of overlap with the specific method he and Alpöge had been steadily advancing, according to BigGo's reporting.^[5]

"Why would you ruin your career?" — OpenAI researcher Sebastien Bubeck, according to Buckmaster's account of a call in which Bubeck reportedly proposed excluding Alpöge from credit because of his employment at Anthropic, per BigGo Finance's reporting^[6]

The Specific Credit Dispute That Turned This Into a Story About Ethics, Not Just Math

The detail that has drawn the sharpest reaction is not the overlap in approach itself, but what Buckmaster says happened when he raised it. According to a separate BigGo Finance report, Buckmaster said OpenAI's Sebastien Bubeck proposed stripping Alpöge of credit specifically because his employment at a rival lab was, in Bubeck's own framing, "a sticking point," reportedly asking Buckmaster directly, "Why would you ruin your career?"^[6] MIT Technology Review's own reporting on the controversy captures the stakes of that specific allegation plainly: if OpenAI's models did train on Buckmaster and Alpöge's unpublished work, or if its agents somehow gained access to it, the company's failure to properly credit the two researchers reflects poorly on it regardless of whether the underlying mathematical result itself is sound.

OpenAI has publicly disputed the characterization of events, according to MIT Technology Review's reporting, while acknowledging, per BigGo Finance's account, that the company cannot fully rule out that de-identified data from researchers' use of its own products may have contributed to improving its models over time.^[7] That acknowledgment, even hedged, is itself significant, since it points to a structural transparency problem that extends well beyond this single dispute: researchers using a frontier lab's own tools to develop unpublished work have limited visibility into whether or how that usage data feeds back into the same company's future model training.

Two Proofs, Neither One Independently Verified

A detail that has received less attention than the credit dispute itself, but matters enormously for how seriously either claim should currently be taken, is that neither side's proof has cleared outside review. According to Rolling Out's reporting, Buckmaster and Alpöge's own most advanced result, the one closest to the actual Millennium Prize problem, remains unpublished pending a final Lean verification check, while OpenAI's 100-page proof has not been reviewed by anyone outside the company. Two competing mathematical claims are currently circulating publicly, and neither has cleared independent scrutiny, a detail that should temper how definitively any outlet, including this one, treats either result as a confirmed mathematical breakthrough at this stage.

What is not in dispute is the underlying quality of Buckmaster and Alpöge's work on its own terms. Fields Medalist Terence Tao described their approach as "a remarkable achievement," an assessment Rolling Out's reporting noted is entirely separate from, and unaffected by, the dispute over who deserves credit for what followed.

Why This Is the Second Story Like This in a Single Week

This controversy did not occur in isolation, and its timing against a separate announcement makes the underlying pattern harder to dismiss as a one-off. Rolling Out's reporting notes this is the second such story in a week involving the same two labs, following Anthropic's own announcement that Claude had produced a computer-checked proof of Fermat's Last Theorem. Taken together, both stories point toward the same emerging dynamic: frontier AI labs are now racing each other directly on open mathematical problems, at a pace and resource scale, thousands of coordinating agents, millions of dollars in compute, that individual academic researchers simply cannot match, regardless of how far along their own unpublished work already is.

BigGo Finance's reporting adds a further layer of context on OpenAI's internal model itself: on a curated set of open math problems, the model achieved a 25 to 45% solve rate, compared to GPT-6 Astra's 10 to 15%, a jump that arrives just days after Depth Grid covered Astra's own launch and Greg Brockman's declaration of "the AGI era." Reading the two announcements together suggests OpenAI already has a materially more capable system running internally than the model it released publicly just days earlier, a detail with its own separate implications for how quickly capability is advancing relative to what the public actually has access to at any given moment.

The Uncomfortable Question This Episode Raises

Axios' own analysis of the situation, offered by reporter Ryan Heath, framed the structural tension at the center of this story with unusual directness: it is striking that OpenAI spent millions of dollars in compute chasing a math breakthrough specifically after hearing that outside researchers, including an employee of a direct rival, were close to one themselves, raising the question of what happens when the company providing scientists with the very AI tools they use for research can also mobilize vastly greater resources to compete directly against them once it learns what they are working on. That question extends well beyond this single mathematics dispute, and directly touches the broader trust relationship between frontier AI labs and the researchers, in any field, who use their tools to pursue original, unpublished work.

What to Watch Next

The most consequential near-term signal will be independent mathematical review of both proofs, since neither OpenAI's 100-page result nor Buckmaster and Alpöge's own advanced result has cleared outside verification as of this writing. Beyond the mathematics itself, watch for whether OpenAI issues any further public statement addressing the specific credit allegations directly, and whether the Clay Mathematics Institute itself weighs in more formally on how it will evaluate competing claims to a problem two separate parties now say they have solved.