Has Navier-Stokes Been Solved? OpenAI's Announcement, Explained Carefully

Ten thousand AI agents, 88 hours, and several million dollars in compute. Here's what's actually been proven, and what hasn't.

⚠ Developing story. This article reflects the state of the announcement as of September 15, 2026. The result has not been verified by the Clay Mathematics Institute or by independent peer review.

On September 8, 2026, OpenAI announced it had obtained a formal result on one of the seven Millennium Prize Problems: Navier-Stokes existence and smoothness. The headline that circulated across social media and mainstream outlets was, in many cases, simply "AI solves a Millennium Prize problem." The reality, as usual, is considerably more nuanced — and no less interesting for it.

What the Navier-Stokes problem actually asks

The Navier-Stokes equations describe how fluids move: the air around a plane, water in a pipe, blood in an artery. They're used constantly in engineering and meteorology because, in practice, they work extremely well. The Millennium Prize Problem asks something different, and much deeper: can it be proven mathematically that, starting from smooth, reasonable initial conditions, solutions to these equations always exist and remain "smooth" forever (in the unforced case — that is, with no artificial external forces applied to the fluid), or can the opposite be proven: that at some point, in finite time, a singularity appears — a kind of mathematical infinity — even though the starting conditions were perfectly reasonable?

Nobody has known this with mathematical certainty for over a century, even though the equations are used daily with no practical issue. That contradiction — they work perfectly in practice, yet we can't prove they "should" always work — is part of what makes the problem fascinating.

What OpenAI actually announced

According to coverage from specialist outlets such as Quanta Magazine, a system of around 10,000 artificial intelligence agents worked in a coordinated fashion for roughly 88 hours, exchanging close to five million messages with each other, at an estimated cost of several million dollars in compute. The result: a proof that singularities can appear in finite time in a forced variant of the three-dimensional Navier-Stokes equations, formally verified with the Lean proof assistant.

The fact that the proof is "verified in Lean" matters: it means a computer system has checked, step by step and unambiguously, that the chain of logical deductions is valid within the specified formal rules. That's a much higher bar of rigour than a traditional proof reviewed only by humans.

"OpenAI itself has been explicit: it isn't claiming the Clay Institute's million-dollar prize for this result, because what's been proven is a version of the problem — not the exact problem."

The distinction that changes everything: forced vs. unforced

This distinction isn't a minor technicality: it's exactly the difference between "we've proven this kind of extreme behaviour is possible in a controlled scenario" and "we've proven it happens in the natural scenario the original problem describes." Both are mathematically interesting, but only the second would resolve the Millennium Prize Problem as it's actually stated.

A dispute in the background

The announcement wasn't free of controversy. According to press coverage, mathematician Tristan Buckmaster of New York University had been working on a related result — together with Levent Alpöge, he had previously obtained a Lean-verified proof for the Euler equations, and an unverified proof for a "somewhat easier" version of Navier-Stokes. Buckmaster suggested that the momentum behind OpenAI's project was driven by rumours about his own unpublished research in collaboration with Anthropic, something an OpenAI representative partially acknowledged. It's a reminder that even in pure mathematics, priority and credit for work can generate friction just as intense as in any other competitive field.

Mathematicians such as Charles Fefferman of Princeton — a recognised authority on this specific problem — have reacted with enthusiasm to the technical advance, crediting much of the conceptual groundwork to earlier analytic work by mathematicians Diego Córdoba and Luis Martínez-Zoroa, which both Buckmaster and OpenAI's agents built on.

Why "unverified" doesn't mean "false"

As we explain in our article on the Poincaré conjecture, the Clay Mathematics Institute's official process requires publication in a peer-reviewed journal, a minimum two-year period of scrutiny by the mathematical community, and final validation by an expert committee. Perelman published his proof in 2002-2003 and it wasn't officially certified until 2010 — eight years, with entire teams dedicated solely to checking that every step was correct.

A result being "verified in Lean" greatly reduces the risk of an internal logical error, but it doesn't replace that process: the mathematical community still needs to confirm that the Lean formalisation actually captures what's intended to be proven, not just that the chain of symbols is internally consistent. That's exactly the stage this result is at right now.

What's worth remembering

• OpenAI's result is a real technical advance, with formal Lean verification, on a forced variant of the problem.
• It doesn't resolve the exact Millennium Prize statement (the unforced case), and OpenAI has explicitly acknowledged this.
• The Clay Institute's official verification process hasn't begun, let alone completed.
• There's an unresolved priority dispute involving prior research by other mathematicians.
• The only Millennium Prize Problem officially solved remains, to this day, the Poincaré conjecture.

We'll update this article if the Clay Mathematics Institute makes an official statement, or if a verified proof of the unforced case appears.