OpenAI accused of math problem foul play

OpenAI claims its unreleased frontier model just solved one of mathematics’ most legendary open problems. The problem is, the mathematicians who paved the way say the AI giant effectively stole their lunch money.

On Tuesday, NYU math professor Tristan Buckmaster and Anthropic researcher Levent Alpöge posted three proofs pushing forward the Navier-Stokes existence and smoothness problem, a $1 million Millennium Prize challenge. Almost immediately after, OpenAI dropped its own full proof of the same problem, crediting an unreleased next-generation model that ran thousands of agents around the clock.

Now, a bitter fight over drafts, credit, and academic intimidation is overshadowing one of the biggest mathematical milestones of the century.

10,000 agents and $22.5 million in compute

The Navier-Stokes equations describe how fluids flow, but mathematicians have spent decades trying to prove whether smooth solutions always exist. To crack it, OpenAI deployed an internal model described as significantly more capable than the newly launched GPT-6 Astra.

According to OpenAI, roughly 10,000 agents ground away simultaneously for 88 hours, burning through 300 billion output tokens. That amounts to about $22.5 million worth of compute at Astra rates. Sam Altman called the breakthrough one of the most amazing moments in OpenAI’s history, and the company told The New York Times it has already made substantial progress on a second Millennium Prize problem.

Allegations of leaks and career threats

Buckmaster isn’t celebrating. He claims OpenAI launched its sprint only after hearing about their private progress, choosing a highly unusual route through smooth forces that almost nobody else was pursuing.

Buckmaster and Alpöge spent a year using OpenAI’s Codex to assemble their work. Buckmaster suspects their drafts may have leaked or fed back into training runs. OpenAI confirmed its sprint started on September 1 based on rumors that two Millennium problems had fallen, but insisted no user data was accessed to generate the proof, while admitting it cannot entirely rule out that de-identified product usage improved its underlying models.

Things turned personal when OpenAI mathematician Sébastien Bubeck allegedly pressured Buckmaster to strip Alpöge of credit due to his ties to rival Anthropic. When Buckmaster pushed to go public, Bubeck allegedly asked why he would ruin his career and warned that he didn't have to be nice. Bubeck has denied the claims, calling them false and inflammatory.

Why it matters

If you are an academic or proprietary researcher pasting cutting-edge drafts into commercial coding models, this is a massive warning sign. Regardless of opt-outs or privacy clauses, the suspicion that frontier labs can leverage user workflows to identify high-potential breakthroughs creates an enormous chilling effect. Feeding unfinished work into proprietary systems right now carries a real risk of getting front-run by sheer infrastructure.

OpenAI showed how brute compute can rewrite scientific discovery, but its tactics look troubling. Unleashing millions of dollars in compute to beat independent researchers to the punch after catching wind of their methodology smacks of industrial bullying rather than pure scientific triumph. In our view, turning frontier math into an aggressive corporate landgrab risks alienating the very researchers AI labs need to win over.