The answer arrived about 88 hours after the first agents started running. OpenAI said on Tuesday that a swarm of roughly 10,000 coordinated AI agents had solved a version of the Navier-Stokes existence problem, one of the seven Millennium Prize problems in mathematics. The equations describe how fluids such as air, water, and blood move, and a rigorous solution has eluded mathematicians since the problem was posed in 2000.
The Navier-Stokes problem is one of seven set out by the Clay Mathematics Institute in 2000, each with a one-million-dollar prize. Only one, the Poincare conjecture, has been solved; Grigori Perelman proved it in 2003 and declined the money. The prize rules require publication in a peer-reviewed journal and a two-year waiting period before any award, hurdles no AI-produced result has yet cleared.
The company revealed something else alongside the result: the model behind it is an internal system “significantly more capable” than GPT-6 Astra, the model it has shown the public. On a set of selected open math problems, the internal model reached a pass rate of nearly 48 percent, almost three times the best result OpenAI attributed to Astra. The compute bill ran into the millions of dollars, the company said.
The result, if it holds, is more than a benchmark. OpenAI said the research suggests three-dimensional fluid motion can develop a singularity in finite time, a point at which the equations break down in describing the fluid as a continuous medium. That finding, if confirmed, would answer a question that has framed one of the hardest problems in analysis for a quarter century.
Not everyone is convinced the breakthrough was independently earned. Tristan Buckmaster, a mathematician at New York University, and Levent Alpoge, a researcher at Anthropic, had published their own AI-assisted study of related fluid dynamics equations shortly before OpenAI’s announcement. Buckmaster questioned whether OpenAI turned to the problem only after hearing of their work, and suggested the company adopted a rare research method the pair had spent months developing. OpenAI said its team had no access to the researchers’ work before it was published, had not touched any specific user data, and used proof methods that differ significantly from theirs.
Independent verification is the harder test. Mathematicians will not accept a result because a company says a model found it; they will want a proof they can read and check, possibly machine-verified in a system such as Lean, where every step is checked by software rather than human judgment. OpenAI has not said whether its agents produced such a proof or something closer to a claim awaiting confirmation. Until that distinction is settled, the community’s default posture will be the one Buckmaster has taken: skepticism.
The dispute cuts to a tension that has grown as AI labs push into research territory long owned by universities. When a company’s model produces a mathematical result, the question of priority and method is hard to settle, because the model’s reasoning path is not fully legible even to its creators. Buckmaster’s challenge is an early test of how the mathematics community will assign credit for results produced at machine scale.
The claim fits a wave of recent machine progress in mathematics. DeepMind’s AlphaProof reached a silver-medal level at the 2024 International Mathematical Olympiad, and OpenAI’s earlier reasoning models have scored well on competition problems and on FrontierMath, a benchmark whose questions tax even specialists. What makes the Navier-Stokes claim different is its ambition: a famous open problem, not a benchmark, with a result that overturns assumptions about how fluids behave.
OpenAI’s framing is that the achievement belongs to its infrastructure as much as any single insight. Ten thousand agents working in parallel, a model far beyond what it has released, and a compute spend measured in millions of dollars describe an industrial approach to a problem that has historically rewarded individual brilliance over years. Whether that approach produces a proof that survives peer review is the open question.
The company has not published the full solution for independent verification, and the Navier-Stokes problem carries a one-million-dollar prize from the Clay Mathematics Institute, which requires publication in a peer-reviewed journal and two years of scrutiny before any award. OpenAI has said nothing about whether it will pursue that process. For now, the claim rests on the company’s own word and a chart of pass rates.
The episode is also a preview of what internal models mean for the industry. OpenAI’s most capable system is not the one its customers can use. The gap between what the company builds for itself and what it sells is widening, and this week offered a glimpse of how large that gap has become.
