OpenAI says it cracked Navier-Stokes, one of math's grand challenges.

OpenAI says it has solved one of math’s marquee problems: the Navier-Stokes challenge from the Clay Mathematics Institute’s Millennium Prize list. The company says a multi-agent system, coordinated by an unreleased internal model and at one point spread across 10,000 sub-agents, proved that there are conditions under which the equations break down.

That is a big claim on its own. Navier-Stokes sits at the awkward intersection of physics and pure mathematics, where the equations work beautifully in practice but still resist a full proof about whether singularities can form under all conditions. It matters for things like fluid dynamics, weather forecasting, and aircraft design. It is also one of those problems that has become a symbol of mathematical stubbornness.

But the route OpenAI took is already causing trouble. Tristan Buckmaster of NYU’s Courant Institute says he and Anthropic’s Levent Alpöge, using AI tools from both companies, independently reached an almost identical line of attack on part of the problem after months of work. He says they built on earlier ideas from Diego Cordoba and Luis Martinez-Zoroa, and that progress only sped up once they brought in the models.

Then came the accusation that turned this from an impressive result into a public mess. Buckmaster says OpenAI pressed him for a call, then appeared to have arrived at the same approach after rumors spread that Anthropic was close to a breakthrough. He also says he was asked to remove Alpöge’s name from a paper if he wanted to publish alongside OpenAI’s claim, and that OpenAI researcher Sebastien Bubeck then threatened him. Bubeck denies that OpenAI used their prompts or proofs, and says the company did not see their work before it was public.

OpenAI also says the effort was expensive in compute. In a briefing, it said the project used at least 1,000 times the computing resources it had spent on some earlier math problems, which it said cost about $2,000 in compute. So the bill was roughly in the $2 million range. That is a lot of machinery for a result that now comes with an asterisk and a lawyerly amount of bad blood.

The deeper complaint here is familiar: AI companies love a clean headline, while mathematicians care about the path, the dead ends, and the ideas that survive the failure. Terrence Tao has been arguing exactly that. If the answer is all that matters, the people doing the work end up looking like props in their own field.