For 165 years, the Navier-Stokes equations have described how fluids move, water in a pipe, air over a wing, blood through a vein, without anyone being able to prove whether those equations could ever break down. It is one of seven Millennium Prize Problems, the set of century-defining open questions the Clay Mathematics Institute put a $1 million bounty on back in 2000. Only one has ever been solved, Grigori Perelman’s proof of the Poincare conjecture in 2003. On September 8, OpenAI said an internal, unreleased model had just solved a second one. Almost immediately, two mathematicians said the company had cut in line to do it.
OpenAI’s claim is specific: a research system the company describes as “significantly more capable” than its current GPT-6 Astra model produced a proof that the three-dimensional Navier-Stokes equations can develop a singularity, a point where the math itself stops making physical sense, in finite time. That is the exact question the Clay Institute has been offering $1 million to answer since 2000. OpenAI says it does not plan to collect.
Quick facts
- OpenAI says an unreleased model proved 3D Navier-Stokes equations can form a singularity in finite time
- The effort began September 1 and produced a result within roughly 88 hours of dedicated compute
- The full run used up to 10,000 AI agents working in parallel and around 300 billion output tokens
- OpenAI estimates the compute cost in the millions of dollars, and is not seeking the $1 million prize
- No independent peer review or Clay Institute certification has happened yet
- NYU mathematician Tristan Buckmaster and Anthropic-affiliated researcher Levent Alpöge announced related results about 12 hours earlier
A week that started with a rumor
According to OpenAI’s own account, the project did not begin as a planned assault on a famous problem. It began with gossip. The company says its researchers heard, in the final days of August, that two of the seven Millennium Prize problems might already be close to falling to outside researchers using AI-assisted methods. Sam Altman has since confirmed the specific trigger: OpenAI moved after hearing rumors that models built by Anthropic, its closest rival, had made serious progress on a major open problem.
OpenAI pointed its newest, still-unreleased internal model at the remaining Millennium problems starting September 1. Over the following week, the system consumed an estimated 300 billion output tokens, worth roughly $22.5 million at the company’s own published API rates for its current flagship model, coordinating as many as 10,000 agent instances working on different pieces of the problem at once. OpenAI says the core proof itself came together in about 88 hours of that window. It is the kind of number that would have been unthinkable as a research budget for a single unpublished result even two years ago, and OpenAI is treating the spending itself as part of the story, evidence of how much capability money can now buy on demand.
The 12 hours that changed the story
The trouble is what happened just before OpenAI published. Tristan Buckmaster, a mathematician at NYU who has spent years working on fluid singularity problems, and Levent Alpöge, a researcher affiliated with Anthropic, announced their own results roughly 12 hours ahead of OpenAI’s post. Their announcement covered several problems closely related to full Navier-Stokes regularity, the product of weeks of work using AI tools adapted from both OpenAI’s and Anthropic’s models, not a single week-long sprint.
“OpenAI fought dirty on career-making math problem.”
Characterization of Tristan Buckmaster’s objections, as reported by TechCrunch
Buckmaster’s objection is not really about who typed the final line of the proof. It is about timing and framing. By his account, OpenAI’s own explanation for why it launched the September 1 effort in the first place was the rumor of his and Alpöge’s unpublished progress. Racing an unannounced academic result to publication, using a commercial model trained in part on the broader mathematical literature both teams were drawing from, then presenting the outcome as a clean OpenAI milestone, is a very different story than an independent lab quietly cracking a 165-year-old problem on its own timeline. Whether the two results even resolve the identical open question, full global regularity versus a set of closely related singularity results, is itself part of the dispute, and it is the kind of distinction that normally gets sorted out in slow, unglamorous peer review rather than dueling press releases 12 hours apart.
What actually counts as proof here
Winning the Clay Institute’s money was never the point, and OpenAI has been explicit that it is not pursuing the prize. That is a meaningful admission on its own, because the Millennium Prize rules require a result to be published in a qualifying journal and survive two years of open scrutiny before any money changes hands, regardless of who or what produced it. OpenAI’s proof has not gone through that process. Neither, for that matter, has Buckmaster and Alpoge’s related work. Both results currently exist in the same unsettled state: publicly announced, widely reported, and not yet independently verified by the mathematicians whose job it is to check this kind of thing line by line.
| Detail | What is known |
|---|---|
| Problem claimed solved | 3D Navier-Stokes global regularity, one of 7 Millennium Prize problems |
| Compute used | ~300 billion output tokens, ~10,000 parallel agents |
| Estimated cost | Millions of dollars, roughly $22.5 million by published API rates |
| Core solve time | ~88 hours, inside a roughly week-long project |
| Prize claim | OpenAI says it is not seeking the $1 million Clay Institute prize |
| Independent verification | Not yet completed for either team’s results |
The Navier-Stokes equations describe exactly this kind of turbulent motion, and mathematicians have never been able to prove the equations always stay well-behaved. Photo via Pexels.
Not the first time a lab has raced a proof to market
The pattern here is starting to look familiar. Anthropic said in early September that its Claude models had produced the first fully machine-checked proof of Fermat’s Last Theorem, a 13-million-line formalization that took 11 days and drew its own scrutiny from the formal mathematics community before independent mathematicians signed off on it. OpenAI has also been pushing hard on the idea that its models can do original research work with minimal supervision, an argument it made just a day before the Navier-Stokes announcement when it described a system it called an automated research intern, a claim that likewise drew criticism for OpenAI writing both the test and grading its own results. Frontier labs are increasingly treating unresolved mathematics as a public leaderboard, and leaderboards reward whoever posts first, not necessarily whoever worked longest or most carefully.
That dynamic cuts against the normal incentives of mathematics, where credit has traditionally gone to whoever’s proof survives scrutiny, not whoever’s press release goes out first. If two teams are drawing on similar underlying ideas, sharpened by AI tools trained on much of the same published literature, disputes over who “really” solved a problem are likely to keep recurring as more labs point increasingly capable models at famous open questions.
What happens now
Nothing here is settled. OpenAI has not released the model responsible for the proof, has not said when independent mathematicians will get to examine the full argument, and has not addressed Buckmaster’s specific objections beyond the original framing of its announcement. The Clay Mathematics Institute, for its part, has given no indication that it is accelerating its normal review timeline for anyone. Until that review happens, both the OpenAI result and the Buckmaster-Alpoge work remain claims rather than confirmed mathematics, however many billions of tokens or however many years of specialist effort went into producing them. The most likely near-term outcome is not a resolution but a long, technical argument playing out in preprints and conference talks, the same way most disputed mathematics eventually gets settled, just with a much bigger compute bill attached this time.
