OpenAI’s Navier-Stokes Claim: Did AI Just Solve a Millennium Problem, or Not Quite?
In a nutshell
everything on the web starts with the domain
This is, on the face of it, one of the most extraordinary claims ever made about artificial intelligence: that on 8 September, roughly ten thousand AI agents, working for eighty-eight hours, produced a proof for one of the seven Millennium Prize Problems — mathematical questions so hard that only one has ever been solved. If true, it would mark the moment AI moved from assisting research to generating it at the deepest level. The capability on display is genuinely staggering. Whether "solved" is the right word is a different question entirely, and it is the one worth getting right while the rest of the world reaches for the headline.
What OpenAI Announced
The facts, as OpenAI presented them, are remarkable and worth stating precisely. An internal, unreleased model — one the company describes as significantly more capable than its just-launched GPT-6 Astra, and still in training — was set loose on the Navier-Stokes existence and smoothness problem, which concerns whether the equations governing fluid motion can break down into a singularity. At its peak the effort ran around ten thousand AI agents in parallel, exchanging 2.7 million messages and generating some 130 billion tokens over 88 hours, at a compute cost running into the millions of dollars. The agents produced a proof of a finite-time blowup: a vortex that tightens and spins until its velocity becomes infinite in finite time, while the fluid's total energy stays bounded. GPT-6 Astra then spent 17 hours formalising and checking the argument in Lean, a language that verifies each logical step. A smaller group of about a hundred agents separately produced a disproof of an Euler regularity question. OpenAI published a 165-page proof and the Lean files for anyone to inspect. On raw capability, this is a genuine leap, and it deserves to be called that.
What "Solved" Actually Means Here
Now the precision the moment demands, because four facts change the story from "solved" to "claimed." First, the Clay Mathematics Institute, which awards the Millennium Prizes, still lists Navier-Stokes as unsolved — and it requires any proposed solution to survive prolonged scrutiny and win broad acceptance before it counts. Second, OpenAI itself is not claiming the million-dollar prize; it released the work as evidence of capability, not as a prize submission. When the company that did it declines to call it a solution, that is a signal worth heeding. Third, this is a proposed proof, 165 pages of it, that human mathematicians now have to inspect line by line — a process that has historically taken months or years, and that has sunk many confident claims before. Fourth, and most immediately: the fight has already started.
The Fight Already Starting
Within a day of the announcement, mathematicians and rivals began contesting OpenAI's account — not only whether the proof holds, but how independently the AI actually got there. By OpenAI's own admission, the work was inspired by rumours that two human mathematicians, Alpöge and Buckmaster, had a proof of blowup for a somewhat easier version of the problem. That raises a genuine and unresolved question: did the agents forge a path, or follow a trail humans had already blazed? This matters enormously for the claim being made. "AI independently solved a Millennium Problem" and "AI completed a proof along lines human mathematicians had already sketched" are very different sentences, and which one is true is exactly what the coming scrutiny must decide. The Lean formalisation is a real strength here — it means the logic can be machine-checked — but formal validity is not the same as resolving the Millennium Problem itself, because everything depends on whether the setup faithfully captures the actual question.
Claim and Counter-Claim
The case for excitement is strong and shouldn't be diluted by caution into dismissal. Even setting aside the "solved" label, an AI system producing a 165-page, machine-formalised proof at the frontier of mathematics is an achievement that would have seemed like science fiction two years ago. If it holds up — and it may — it genuinely marks a transition from AI-assisted to AI-generated frontier research, and that is historic regardless of the prize.
The case for restraint is equally strong and rests on the record, not on skepticism for its own sake. Extraordinary mathematical claims collapse under scrutiny more often than they survive it; the Clay Institute's caution exists precisely because history is littered with confident proofs that turned out to be flawed. The independence question is unresolved and serious. And the compute-brute-force nature of the approach — ten thousand agents, millions of dollars, 130 billion tokens — raises a question the celebration skips: is this deep mathematical insight, or an industrial search that happened to land on an answer humans were already near? The honest synthesis is the one OpenAI itself implicitly adopted by not claiming the prize: this is a stunning demonstration of capability and a proposed proof under dispute, not a settled solution. Be amazed. Wait for the mathematicians.
The European Perspective
Look past the proof to the method, because that is where the story reaches everyone, Europe included. For all of history, a mathematical breakthrough could come from anywhere — a patent clerk in Bern, a reclusive genius in St Petersburg. The only Millennium Problem solved before this was the Poincaré conjecture, cracked by one man, Grigori Perelman, working largely alone, who then refused the million dollars. That is the old model of discovery: insight, not infrastructure, and available to any mind anywhere. What OpenAI has demonstrated, whether or not this specific proof holds, is a new model — discovery as a function of compute, where the breakthrough goes to whoever can afford ten thousand agents and millions of dollars of processing. And that should concern anyone outside the two AI superpowers, Europe most of all.
If frontier science becomes something you buy rather than something you think, then the same concentration we have tracked in models and cloud extends to knowledge itself: the great discoveries flow to those with the largest data centres, and a continent that cannot match that compute does not just import its AI — it imports its science.
There is a crucial counterweight, and it is the most European note in the whole story. The arbiter of whether this proof is true is not OpenAI, and not its compute budget. It is the global community of human mathematicians, working in the open, checking the argument line by line, refusing to accept a claim until it earns acceptance — a tradition centred as much in Europe's universities as anywhere on earth. In an age where machines can generate proofs faster than humans can read them, that patient, distributed, human authority over what counts as true becomes more precious, not less. The machine may have written the proof. But it is still people who decide whether it is real — and that, for now, is a form of sovereignty no compute budget can buy.
We are not first. We are right.