OpenAI says it cracked the Navier-Stokes Millennium Problem: 10,000 AI agents solved a 90-year math mystery in 88 hours

Let's unpack the spectacle of 'AI solving math.' BBC's reporting reconstructs the full timeline: in late August, OpenAI began training a new internal model and quickly found it excelled at mathematics. On September 1, OpenAI heard rumors that 'two Millennium Prize problems had been resolved,' so it decided to put thousands of AI agents trained on the new model to work on the remaining problems. By September 5 — roughly 88 hours after setting 10,000 agents on the task — OpenAI had a solution to the Navier-Stokes existence and smoothness problem. At its heart, the question asks: do smooth solutions of the Navier-Stokes equations, which describe three-dimensional fluid motion, lose their smoothness in finite time (a so-called 'blow-up')? This is the mathematical bedrock of turbulence, a phenomenon still among the least understood in physics. The engineering details OpenAI disclosed are staggering: the agents exchanged nearly 3 million messages and consumed 130 billion output tokens on Navier-Stokes alone — at OpenAI's own pricing for output from its most advanced models, that effort would have cost roughly $10 million (about £7.3 million).

However, there is a long road between 'solved' and 'accepted.' OpenAI itself is cautious: it says the result resolves two of the four statements the Millennium Prize demanded, and the company has no intention of claiming the $1 million prize. The Clay Mathematics Institute has not yet independently verified or publicly accepted the solution — under the Millennium Prize rules, any solution must pass rigorous peer review before the award is granted, and since OpenAI used an unreleased internal model and has not published full details of the proof, outside mathematicians cannot yet check it line by line. Even so, the symbolic weight is enormous: The New York Times called it 'the most dramatic sign yet that artificial intelligence is fundamentally transforming higher mathematics.' Over the past year, AI systems have solved a wide range of problems that bedeviled mathematicians for decades, but none approached the complexity or depth of Navier-Stokes. OpenAI treated its ~10,000 agents as a 'research team' — exchanging messages, dividing verification work, iterating on reasoning — and this pattern of an 'agent swarm conquering scientific problems' excites researchers even more than a single model's capability leap.

What turned this from a tech story into an academic drama is the fierce dispute over credit. NYU mathematician Tristan Buckmaster issued a public statement on September 8: he and Anthropic researcher Levent Alpöge — an employee of OpenAI's direct rival — had been using AI tools (including OpenAI's Codex) to study a related version of the Navier-Stokes problem — the friction-free version of the equations — and had already achieved a verified result showing a singularity can develop in finite time. Buckmaster claims that on September 3 he learned that 'information about our progress had been passed to OpenAI'; that OpenAI urgently requested a phone call starting September 3; and that during a call on September 6 he learned OpenAI was about to claim its unreleased model had solved Navier-Stokes using exactly the same line of attack he and Alpöge had used. The Information and other outlets followed up, reporting that OpenAI researchers allegedly pressured Buckmaster during the call to drop an Anthropic co-author from the related paper. Fields Medalist Terence Tao, meanwhile, called Buckmaster and Alpöge's fluid-equation proofs 'a remarkable achievement,' saying the work could help solve Navier-Stokes. For now the two accounts contradict each other: Buckmaster and Alpöge's result is public and verifiable, while OpenAI's full proof has not been publicly disclosed — who moved first, and whether there was any 'borrowing,' cannot be settled from the outside yet.

Zooming out, whatever the final verdict, this marks a milestone: AI has touched one of the most dazzling jewels in mathematics' crown. Over the past two years, AI's role in math has evolved from a 'calculator' to a 'conjecture generator' (DeepMind's FunSearch discovered new structures in number theory and combinatorics), and now to an 'autonomous proof assault team' — what OpenAI demonstrated is no longer a single model answering questions, but ten thousand agents collaborating like a research institute: proposing conjectures, dividing verification, cross-checking, iterating. This is a double-edged sword for mathematics itself: optimists believe AI can free mathematicians from tedious derivations so they focus on true insight — Tao himself is an active practitioner of AI-assisted proof; conservatives worry that black-box proofs cannot be understood by humans and that 'brute-force search' proofs will dilute the beauty and rigor of the mathematical tradition. For ordinary readers, the more practical takeaway is this: when OpenAI is willing to burn roughly $10 million of compute on a proof it won't even claim a prize for, it shows frontier labs are investing in 'agent swarms solving real-world problems' at almost any cost — mathematics is just the first castle to fall; medicine, materials and climate, which equally depend on long-chain reasoning, are probably next.

As I write this, I recall a classic saying in mathematics: a hard problem does not disappear because you announce a solution — it is only truly solved after it is verified. OpenAI's Navier-Stokes claim essentially thrust an experiment into the spotlight: can AI reach the deepest recesses of human intellect? Ten thousand agents, 88 hours, 130 billion tokens, $10 million of compute — the payoff may be a proof that transforms fluid dynamics and mathematical research paradigms, or a scholarly dispute that takes years to untangle. But regardless of the outcome, one thing has changed: when a frontier lab is willing to attack a pure-mathematics problem at this scale, 'AI scientist' is no longer a metaphor — it is an engineering reality taking shape. For ordinary people following AI, the news worth remembering is neither 'solved' nor 'controversial,' but this: humanity is now seriously handing its hardest thinking to tireless algorithmic agents.