On 8th September 2026, OpenAI announced that frontier models incorporating ten thousand agents had cracked one of the seven Millennium Prize Problems, the most famous open questions in mathematics, each carrying a $1m bounty. Headlines declared that a machine had finally done the unthinkable and won the money. Interestingly, OpenAI is not asking for the prize, and the Clay Mathematics Institute, which sets it, still lists the problem as unsolved.

The problem is the Navier-Stokes equations, named for Claude-Louis Navier, a French engineer who first wrote them down in 1822, and George Gabriel Stokes, the Cambridge physicist who two decades later put them on rigorous footing. Navier reached the right equations by the wrong reasoning, from a flawed picture of how molecules jostle; his own suspension bridge across the Seine, the Pont des Invalides, was pulled down before it ever opened. What he left behind has outlasted the bridge. For two centuries the equations have defied a basic question about their own reliability, beating the finest analysts alive, because they are nonlinear. The prize that now hangs on them descends from a piece of theatre: in 1900, at a congress in Paris, David Hilbert set mathematics 23 problems for the coming century. In the year of 2000, in the same city, Landon Clay, an American financier who made his fortune backing fund managers, endowed an institute that chose seven fresh mathematical proof problems and attached $1m to each - behold the Millennium Prize problems. The official Navier-Stokes statement fell to Charles Fefferman of Princeton, a prodigy who had won a full professorship at 22 and a Fields Medal by 29 - who wrote the Navier-Stokes question.

The equations describe how fluids move: water in a pipe, air over a wing, even blood in an artery. Engineers use them daily to design aircraft and forecast the weather. The issue at hand is the proof in complex systems, a small disturbance even in a water pipe can amplify itself without predictable limits, and no one has proved it never will. The fear is a "singularity", a point at which the fluid's speed races to infinity and the mathematics collapses. Prove that ordinary water never does this, or find a case where it does, and the money is yours.

Only one of the seven problems so far has fallen. In 2003 Grigori Perelman, the reclusive Russian mathematics rockstar, proved the Poincaré conjecture, then turned down both the $1m and the Fields Medal, wanting no part of the fuss. The Poincaré conjecture is a famous theorem in mathematics stating that a three-dimensional space without any holes, where every loop can be shrunk to a single point, is topologically equivalent to a three-dimensional sphere. Then, the remaining five, among them the famously sought Riemann hypothesis on the distribution of primes and P versus NP, which asks whether every answer easy to check is also easy to find, are still open. Most carry the accumulated wreckage of failed assaults.

What OpenAI's model actually did is narrower than the headlines would have you believe. It produced a proof that a singularity can form, but only in a "forced" version of the equations, in which an outside hand pumps energy into the fluid on a chosen schedule. Clay's statement has four parts; the forced case, options C and D, is not the one mathematicians prize. Nor did the machine work alone: the scaffolding came from Diego Córdoba and Luis Martínez-Zoroa, two mathematicians in Madrid who spent years building the "cascade" techniques the model agents used. "The heroes of the story", Dr Fefferman said, "are Córdoba and Martínez-Zoroa." Sébastien Bubeck of OpenAI ran some 10,000 agents in parallel for 88 hours, exchanging nearly 5m messages, at a cost Mark Chen of OpenAI put in 16 million dollars. Considering the Millennium Prizes cash value, it might really be glory that Sam Altman is seeking. A computer then checked the argument in Lean, a verification language, over another 17 hours.

The check carries its own asterisk. OpenAI published its files, but the review is self-assessed and no outside group has re-run it, as of yet in full. Terence Tao, the whiz-kid Fields medallist who has laboured on the problem for years, called the approach "a plausible route" beset by "enormous technical difficulties", and declined to endorse it. Stan Palasek of Princeton University warned that the technicality, namely "energy lost to viscosity," might swamp the growth the proof needs.

The episode marks a shift in how mathematics gets done. Machines have lately moved from checking proofs to attempting them, winning medals at the International Mathematical Olympiad and spreading through the field as proof assistants. What unsettles mathematicians is the manner of it. Dr Bubeck said his team "lacked research-level expertise in fluid dynamics and could not meaningfully contribute mathematical content," an odd thing to boast of a proof one's company is claiming. Hours before the announcement, Anthropic researcher Levent Alpöge rushed out a related result on the Euler equations, with fellow mathematician Tristan Buckmaster.

Princeton's Dr Buckmaster says his team sped up pipelines after word of their progress "leaked to OpenAI," and he raised a sharper concern: OpenAI's model was still being trained during the effort, and he wanted to know whether unpublished work had fed it. OpenAI conceded it "cannot rule out" that de-identified usage data helped its models, while calling it unlikely. Bubeck disputed Buckmaster's account of their conversations; Sam Altman backed Bubeck. Buckmaster has described one of the circulating manuscripts as something that "can only be described as AI slop."

Hours before OpenAI's announcement, Tristan Buckmaster of Princeton and Levent Alpöge of Anthropic had rushed out a related result on the Euler equations. Buckmaster says.

Clay Institute, for its part, is in no hurry. Its president, Martin Bridson, promised an evaluation "deliberately unhurried" and "absolutely rigorous"; the rules demand publication and two years of scrutiny before any cheque is written. The open questions now pile higher than the answers. Whether the proof survives in its unforced form, the one that actually carries the prize, nobody yet knows.

But the deepest doubt remains cultural: a field built on openness may stop sharing half-finished work, now that a rumour of progress can trigger a rival's 10,000-agent sprint to publish first. Dr Tao thinks that last danger to be the gravest. The robots have gone quiet, mathematicians are still arguing though.

Shoumik Zubyer is a science correspondent, a researcher at the Bangladesh Atomic Energy Commission and the Space and Environment Research Centre.