OpenAI's Navier-Stokes claim awaits Clay review

By VK

Published 6 Oct 2026, 11:28 UTC

Reporting from The Verge, IEEE Spectrum

OpenAI says its AI solved the Navier-Stokes Millennium Prize problem. The Clay Institute has not verified it, and mathematicians are split.

What it means for founders

Treat lab reasoning claims as unverified until a third party signs off. The most famous AI math claim is still under review a month on. If you pick a model for research, finance or legal work on the strength of headline results, test it on your own problems first.
Budget for verification, not just generation. Williamson's worry is that answers are outrunning understanding. In a product, that gap becomes human review time. A model that produces more candidate answers raises the cost of checking them, so price that labor into any research-heavy feature.
Norms are now a platform risk. OpenAI's math push became a reputational problem because of how results were released. Startups selling AI tools into universities and expert teams inherit some of that distrust, and buyers may ask about data provenance the way Thom did.
Watch two signals. Clay's next update on the Navier-Stokes review, and the first batch of results released with the new panel's input. Clean outcomes make AI research tools easier to sell; messy ones slow adoption.

The story

OpenAI said on September 8 that an internal AI model had cracked the Navier-Stokes problem, one of the Clay Mathematics Institute's seven Millennium Prize Problems. Four weeks on, nobody independent has confirmed it, and much of the reaction is about how OpenAI got there.

What OpenAI claims about Navier-Stokes

According to The Verge, OpenAI's blog post credited an unreleased model more capable than GPT-6 Astra, working with 10,000 concurrent agents, and said training on that model began on August 28. The problem asks whether well-behaved solutions to the equations of fluid motion always exist in three dimensions. Each Millennium problem carries a $1 million prize, and the Poincaré conjecture is the lone one settled so far.

The Clay Mathematics Institute responded on September 11 that the problem had "apparently been settled," adding that its evaluation under the prize rules is deliberately unhurried. It named no solver and set no timeline. Until that review ends, this is a claimed solution, not a confirmed one.

Why mathematicians are pushing back

At the Heidelberg Laureate Forum, which opened on September 13, Fields Medalist Jacob Tsimerman said the result looked decisive while admitting he did not know its details, IEEE Spectrum reported. Fellow Fields Medalist Peter Scholze called the labs' pursuit of famous problems a benchmark exercise and a publicity play, and Geordie Williamson said OpenAI had behaved very badly.

The complaints are mostly about conduct:

Racing rivals. NYU's Tristan Buckmaster had been making progress on Navier-Stokes with Anthropic's Levent Alpöge, and OpenAI appears to have pushed late after learning of it. Buckmaster says his correspondence with OpenAI read as coercive; Columbia's Michael Harris says he trusted that account. OpenAI did not respond to IEEE Spectrum.
Training data. Andreas Thom has asked whether mathematicians' own ChatGPT chats helped OpenAI reach a result in his field, which OpenAI admits built heavily on earlier work by Thom and Gábor Kun.
Careers. Early-career researchers at the forum said they now fear being overtaken by anyone who pastes their problem into a model.

In late September OpenAI announced an independent panel of mathematicians to advise AI companies on presenting and releasing results. Its first job is helping coordinate a large backlog of further results that OpenAI attributes to the same unreleased model, as Enki reported.

What we don't know yet

Whether Clay's review will confirm the proof, and when it will report.
How Clay's rules for assigning credit apply to a result produced by a company's model.
How much say the new panel will actually have over OpenAI's next releases.