OpenAI's Navier-Stokes Claim Is About More Than Math

A race condition in mathematics

OpenAI has announced that one of its unreleased AI models found a solution to the Navier-Stokes Millennium Prize problem—one of the seven hardest unsolved questions in mathematics, carrying a million-dollar reward and a reputation for breaking careers.

The announcement lands in a context that has nothing to do with fluid dynamics. On September 1, OpenAI’s researchers heard rumours that Anthropic, their fiercest competitor, was close to solving two Millennium problems. They deployed an in-development model and a swarm of 10,000 AI agents. Within 88 hours, they had a result. Verification took another 17.

This is not how mathematics is supposed to work. But it may be how it works now.

What the problem actually is

The Navier-Stokes equations describe how fluids move—water, air, molten metal, the atmosphere. They are indispensable to engineering and science. But we do not fully understand them.

The core question is simple to state and brutally difficult to resolve: can the equations produce a singularity, a point where fluid velocity becomes infinite in finite time? In the real world, this never happens—you can zoom into water until you see molecules, not infinite speeds. But do the equations themselves allow it? That is the Millennium Prize problem.

OpenAI says it has shown that yes, under certain conditions, a blow-up is possible. The American Mathematical Society called the AI’s contribution “the final steps” of the solution process. Building on work by several human mathematicians, the model churned through millions of dollars in compute over a few days.

A solution to the similarly formidable ABC conjecture ran to more than 500 pages and took six years for the community to fully digest. This one arrived in five days, generated by machines.

The human work underneath

Here is what the press release does not emphasise: OpenAI did not solve this alone. The result rests on foundations laid by human mathematicians. The company’s researchers had been in contact with Tristan Buckmaster of New York University, whose work was central to the problem.

Buckmaster published a statement hours before OpenAI’s announcement. He said he and Anthropic staffer Levent Alpöge had been using OpenAI’s publicly available models in their own research. When Buckmaster asked whether his unpublished work had been used by OpenAI’s model, he received no answer.

Instead, according to Buckmaster, OpenAI offered to collaborate on the result—if he removed Alpöge’s name. Alpöge works at Anthropic. OpenAI denies making this request.

German mathematician Andreas Thom has raised a broader concern: that OpenAI’s models may have consumed unpublished human research and presented it as AI-generated output. The company insists no specific user data was accessed. The question remains unresolved.

Why mathematicians are alarmed

Terence Tao, the US-Australian mathematician who has been sharply critical of certain trends in AI-assisted research, framed the issue clearly. He warned against the “indiscriminate use of powerful solution-extraction tools” that prioritise short-term problem-solving over genuine understanding.

The concern is not just about credit. It is about incentives.

If knowing your research will trigger a rival company to deploy millions in compute and a thousand AI agents to beat you to a solution, what happens to the culture of open mathematics? Researchers share findings, build on each other’s work, and slowly accumulate understanding. That system depends on trust. Corporate speed runs counter to it.

As Tao put it, if AI companies start racing on rumours, the community may simply stop sharing promising work altogether. The foundation stones of the next wave of discoveries would vanish.

Who wins, who loses

OpenAI wins the narrative. It is positioning itself as the company that can do what humans cannot—or at least, what humans cannot do quickly. Both OpenAI and Anthropic are heading toward planned public listings. A Millennium Prize solution, even one still awaiting formal verification, is excellent advertising.

Anthropic loses the framing, at least temporarily. Its staffer Alpöge was arguably closer to the solution than the public realised. The company now looks reactive rather than pioneering.

Mathematicians lose the conditions that make their work possible. Open and reproducible science is not a slogan—it is a practical system. It requires sharing drafts, citing collaborators, and giving credit where it is due. If that system erodes, the cost is borne by everyone who does hard thinking about hard problems.

Buckmaster and Alpöge stand at the centre of this. Their private collaboration appears to have been a source material for OpenAI’s result. Whether that counts as use of proprietary data or simply engagement with the public literature is a legal and ethical question that has not been answered.

What happens next

The Clay Mathematics Institute has a rule: Millennium Prize submissions cannot be awarded for at least two years after publication. The Navier-Stokes problem remains officially unsolved. Whatever OpenAI has produced will undergo scrutiny that no amount of compute power can shortcut.

But the precedent is already being set. AI companies now know that throwing computational resources at a rumour can produce a headline-grabbing result in days. Human researchers know that their unpublished work may fuel that process without consent, without credit, and without recourse.

The mathematical community’s response will determine whether this becomes a one-off embarrassment or a structural shift. If researchers begin withholding work, the consequences will outlast the press cycle. If they continue to share, they will need new protections—legal, institutional, and cultural—to ensure that speed does not become synonymous with ownership.

OpenAI’s announcement is a milestone, whether or not the math holds up. The deeper question is what kind of research ecosystem survives when the race to publish beats the habit of understanding.