Mathematics & AI
OpenAI Says AI Cracked a $1 Million Millennium Math Problem. The Claim Is Already Under Fire
OpenAI says an unreleased internal AI system has solved the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize Problems that have stood for a quarter century as some of the hardest questions in mathematics. CNN carried the claim on Wednesday. Independent mathematicians have not yet treated it as settled, and a credit fight broke out within hours of the announcement.
The Clay Mathematics Institute listed the seven problems in 2000 and offered $1 million for each correct solution. Only one, the Poincaré conjecture, has been officially recognized. OpenAI says it will not claim the prize. The Clay Institute has not commented.
What OpenAI says it proved
The Navier-Stokes equations, developed in the 19th century by Claude-Louis Navier and George Gabriel Stokes, describe how liquids and gases move. Weather models, aircraft design, blood flow, and ocean currents all rest on them. The Millennium question is whether those equations always produce smooth, well-behaved solutions, or whether a fluid that starts out normal can blow up, reaching infinite speed in a finite time.
OpenAI says its system found that blowup. In the company’s account, an initially smooth fluid at rest, under a smooth force and with finite energy, can develop a singularity. OpenAI says that result covers statements C and D in the Clay Institute’s official formulation, not the classic unforced regularity question often labeled A and B. In plain language, the company is not saying every fluid simulation explodes. It is saying the equations can break.
The work was not done by ChatGPT as the public knows it. OpenAI used an internal model it describes as significantly more capable than GPT-6 Astra, which shipped last week. About 10,000 AI agents worked in groups for roughly 88 hours, from September 1 to September 5. They exchanged millions of messages. Formalizing the argument in Lean, a language that lets a computer check each step of a proof, took another 17 hours using Astra. OpenAI published a write-up and the Lean code. That machine-checkable file is the strongest part of the claim, because other researchers can test the logic without taking the company’s word.
Executives put the cost in the millions of dollars. One figure floated at the press conference put a customer running the same job at current rates at about $15 million. An independent estimate from University of Michigan researcher Karthik Duraisamy landed lower, near $6 million at retail rates. Either way, this was industrial-scale mathematics, not a lone professor at a blackboard.
The other team that was already on the problem
The night before OpenAI’s announcement, New York University mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge released their own work on related fluid equations, including Euler, a close cousin of Navier-Stokes. They had been on the problem for about a year, using large language models that included OpenAI’s Codex and Anthropic’s Claude, and building on methods pioneered by Diego Córdoba and Luis Martínez-Zoroa.
OpenAI says it started its own push on September 1 after hearing a rumor that Millennium problems had been solved. It says neither its researchers nor its agents saw Buckmaster and Alpöge’s unpublished work, and that no specific user data was searched to produce the proof. Chief research officer Mark Chen called the allegations disappointing. The company also offered a narrower concession. It “cannot rule out” that de-identified data from the pair’s use of OpenAI products helped improve its models.
OpenAI has a second line of defense. It says its own proof targets a structurally different theorem than the pair’s. Its Euler result covered the unforced case, while Buckmaster and Alpöge worked the forced version, and OpenAI says it recognizes their priority on that forced problem.
Buckmaster has been careful and pointed at the same time. In a public statement he wrote that he has not seen OpenAI’s proof, does not know what the model did, does not know whether his data was used, and is “not accusing anyone of anything.” Other reporting describes a sharper dispute behind that sentence. OpenAI learned the pair were close, raced the problem with a private model the public cannot inspect, and, according to Buckmaster, floated a deal in which he would get sole authorship only if Alpöge’s name came off because Alpöge works at a rival lab. OpenAI researcher Sébastien Bubeck denied that the pair’s work steered the company’s proof. He also apologized for one line Buckmaster attributed to him, “Why would you ruin your career?”, and OpenAI apologized for some of its language, a partial concession that not every part of Buckmaster’s account is in dispute.
Those are allegations, not findings. They still matter. If scientists cannot feed unfinished proofs into commercial AI tools without fearing that a lab will harvest the direction and publish first, the tools become a trap. That is a governance problem, not a math problem.
Has the prize actually been won?
No. A company blog post and a Lean file are not the same thing as Clay Institute recognition. The Institute has a review process. OpenAI is not applying for the million dollars.
“The indiscriminate use of AI is turning the subject into a meaningless production quota game.”
Terence Tao, mathematician, to CNN
Tao, one of the world’s leading mathematicians, compared AI-only solutions to skipping from the first ten minutes of a film to the last ten. The plot may close, he said, but most of the value is gone.
Even if the Lean proof holds, the scientific meaning is narrower than the headline. Weather apps will not stop working tomorrow. The result is a statement about the equations, under specified conditions, with a forced term. Engineers already know real fluids are messier than the ideal math. The historic claim is that a machine produced a proof on a problem that had resisted people for generations.
What this fight is really about
Two stories are running at once.
- AI labs can now throw thousands of agents and millions of dollars at a single theorem and, if the formalization survives, finish work that defined a field.
- The same labs are racing each other so hard that credit, training data, and unpublished drafts have become combat terrain.
OpenAI wants the public to see a milestone, an internal model that solved what human specialists could not close. Buckmaster and Alpöge want the public to see the week before the press release, when a professor and a rival-lab mathematician were already using the company’s own tools on the same mountain. CNN’s post sold the first story. The second is why replies under that post immediately asked whether the company “stole” a professor’s solution. The honest answer, on the evidence available today, is that theft has not been proved, independence has not been proved either, and the Clay Institute has not spoken.
Why this matters far from Silicon Valley
For countries that will never train a 10,000-agent swarm, the lesson is simpler. The deepest questions in science are being settled inside private firms, on models the public cannot run, under rules the public did not write. If the proof is right, it is a genuine advance in understanding fluids. If the process is as messy as the accusations suggest, it is also a warning about who gets to own the next century of discovery.
Disclosure: This report involves Anthropic, the maker of Claude, an AI tool VPN uses in its newsroom, and one of the mathematicians at the center of the dispute is an Anthropic researcher. All quotes and claims were independently verified against CNN, Quanta, Science, Scientific American, CNBC, Fortune and OpenAI’s own write-up before publication. VPN takes no side in the credit dispute.