OpenAI Says It Solved Navier-Stokes in 88 Hours. Mathematics Is Taking Its Time
OpenAI says 10,000 AI agents cracked Navier-Stokes in 88 hours. The fine print is messier: the proof covers the forced version of the problem, no independent mathematician has finished checking it, and a rival team suspects its ideas leaked through its own AI coding assistant. What really happened, and what it means when machines find answers faster than humans can understand them.
📅
✍️ Gianluca
OpenAI Says It Solved Navier-Stokes in 88 Hours. Mathematics Is Taking Its Time
On September 8, 2026, OpenAI announced that a swarm of roughly 10,000 AI agents, running on an internal model the public has never touched, had produced a proof for the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize problems and among the most famous open questions in mathematics. The whole run took about 88 hours. The BBC headline said OpenAI "cracked a 90-year-old maths problem". Within hours, a New York University professor was publicly accusing the company of racing him to the result after learning about his unpublished work, possibly through his own use of an OpenAI coding tool.
Both halves of that paragraph deserve scrutiny. Because the real story is neither "AI solved one of the hardest problems in mathematics" nor "it is all hype". The truth sits in the fine print, and the fine print happens to be where the most interesting questions about AI, trust and human understanding live.
A problem older than the computer
The Navier-Stokes equations describe how fluids move. Claude-Louis Navier wrote the first version in 1822; George Gabriel Stokes refined it in the 1840s. Two centuries later these equations sit inside every weather model, every aircraft simulation, every CGI ocean in a film. Engineers use them daily and they work. What nobody has ever established is whether the mathematics underneath is sound in the strongest sense: given smooth initial conditions in three dimensions, do solutions always exist and stay smooth forever, or can a perfectly well behaved fluid spontaneously concentrate infinite velocity into a single point, an event mathematicians call a singularity or blowup?
The modern form of the question dates to 1934, when Jean Leray constructed his famous weak solutions and left the smoothness question open. That is the ninety year clock the headlines refer to. Olga Ladyzhenskaya settled the two dimensional case in the 1960s; the three dimensional case resisted everyone. In 2000 the Clay Mathematics Institute selected it as one of its seven Millennium Prize problems, each worth one million dollars, and Charles Fefferman wrote the official problem statement: four precise claims, labeled A through D, where proving either existence and smoothness (A or B) or breakdown (C or D) counts as a resolution.
Ninety Years of Partial Answers
The problem did not sit untouched. In 2016 Terence Tao proved that an averaged version of the equations, one that keeps their energy structure but relaxes their geometry, does blow up in finite time, a result that ruled out entire families of proof strategies. In 2019 Tristan Buckmaster and Vlad Vicol showed that weak solutions of the true equations are wildly non unique, exposing how fragile the standard solution theory is. More recently Diego Cordoba and Luis Martinez-Zoroa developed the "forcing" technique, building singularities in fluid systems that are continuously stirred by an external push. And Princeton mathematician Stan Palasek identified a serious obstacle on the other side: without forcing, viscous energy loss appears to overwhelm the known mechanisms for blowup.
In other words, by 2026 the field had a map: blowup with an external force looked reachable, blowup without one looked blocked, and everyone knew which route was which. Keep that map in mind. It is about to matter twice.
What OpenAI actually claims
According to OpenAI, at the end of August it began training a new internal model that quickly showed unusual mathematical ability, significantly beyond anything the company has released. On September 1, after hearing rumours that two Millennium Prize problems had been resolved, it pointed a swarm of agents built on that model at the remaining ones. By September 5 the swarm had produced a proof that the three dimensional Navier-Stokes equations can blow up: a fluid that starts perfectly smooth and at rest, pushed by a smooth external force with finite energy, develops a singularity in finite time. That resolves statements C and D of Fefferman's official formulation, the two breakdown claims, and it resolves the Clay problem in the negative direction: smoothness does not always win.
The run, by the numbers
Roughly 10,000 concurrent agents worked for about 88 hours, exchanging close to 3 million messages and producing around 130 billion output tokens on Navier-Stokes alone. At OpenAI's own list prices for frontier output, that is in the order of 10 million dollars of compute for one proof.
The output is a 166 page manuscript, "Finite Time Blowup for Navier-Stokes", plus a formalization in the Lean 4 proof assistant published on GitHub, built against the Mathlib mathematical library.
OpenAI says it does not intend to claim the one million dollar Millennium Prize. The stated goal of the release is "to report on the substantial progress of our AI models".
The asterisk in the fine print
Here is where the map from earlier matters the first time. The proof uses forcing: the fluid is not left alone, it is continuously stirred by a carefully chosen smooth push, a bit like a propeller that never stops. Fefferman's official statement explicitly permits this route in statements C and D, so by the letter of the Clay problem, this counts. But most working mathematicians, when they picture the Navier-Stokes problem, picture the unforced question: can a fluid blow up on its own? That version remains open, and Palasek's obstacle suggests it is genuinely harder, not just unfashionable.
So the accurate sentence is narrower than the headlines: an AI system appears to have resolved the two forced statements of the official Clay formulation, while the version of the problem most experts consider the heart of the matter is still standing. That is not nothing. It is arguably a historic result. But "cracked the 90 year old maths problem" flattens a distinction that the entire field is built on, and it is worth noticing that the flattening happened in the coverage, not in the mathematics.
Has anyone actually checked it?
This was the question we most wanted answered before writing this article, and the answer as of publication is: no independent authority has validated the result yet. The Clay Institute has not accepted it and its president, Martin Bridson, said the evaluation would be "deliberately unhurried" and "absolutely rigorous". The institute still lists the problem as unsolved. The manuscript has not been through peer review. What does exist, and what separates this from a pure press release, is the Lean formalization: a machine checkable artifact that anyone can download and verify.
A Compiler for Proofs, and Its Limits
Developers already understand the strongest and weakest parts of this setup, because Lean works like a type checker for mathematics. If the Lean project compiles, every deduction in the formalized argument is correct; there is no room for the subtle gap on page 83 that sinks traditional manuscripts. In that narrow sense, this proof is more verifiable than almost any famous proof in history.
But Terence Tao immediately pointed at the boundary of that guarantee: Lean verifies the statement it was given, not whether that statement faithfully encodes Fefferman's problem. The definitions, the function spaces, the energy conditions, the exact shape of the forcing term all have to be checked by humans against the Clay formulation. Any developer who has watched a green test suite validate the wrong specification knows exactly what this failure mode looks like. The compilation is trustworthy; the requirements review is still pending, and it is being done by people, slowly.
The ugly part: who was first?
Now the map matters the second time. On August 15, NYU professor Tristan Buckmaster and Levent Alpoge, a mathematician employed by Anthropic, quietly achieved something remarkable: working with AI tools, they proved finite time blowup for the Euler equations, the frictionless cousin of Navier-Stokes, along with breakthroughs in two further simplified models. Their proofs were verified in Lean too. Tao called the work remarkable. They had deliberately chosen the forced route through statements C and D, the exact path the field's map marked as reachable, and they were preparing their publication.
Hours before OpenAI's announcement, Buckmaster went public with a very different story. He says information about their unpublished progress reached OpenAI in early September; that OpenAI only began its Navier-Stokes push after that; and that the direction its swarm took is "not the direction one arrives at in a few days by giving a model the problem statement". He had used OpenAI's Codex extensively during the research, and he wonders openly whether his own coding sessions fed the competitor that then raced past him. He also describes a phone call in which, he alleges, an OpenAI representative proposed a version of events that dropped Alpoge, who works for OpenAI's main rival, and responded to his objections with "Why would you ruin your career?".
OpenAI denies the substance: it says it saw none of their work before it became public, that no specific user data was accessed, that the proofs differ significantly, and it publicly congratulated the "remarkable" concurrent work. But its statement contains one sentence worth reading twice: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."
Your AI Sessions Are Someone Else's Data
Strip away the personalities and a structural problem remains, one that applies to every developer reading this. Frontier research, and increasingly frontier engineering, now flows through AI assistants operated by companies that are themselves competitors in the race. When the referee also fields a team, even a scrupulously honest referee inherits a conflict of interest, and "we cannot rule it out" becomes the honest ceiling of any privacy guarantee.
The practical lesson is not to abandon the tools. It is to treat an AI assistant session the way you treat any third party service holding your unpublished work: know the data retention policy of your tier, know what opt outs exist, and assume that anything genuinely novel you type into a competitor's product is, at minimum, a calculated risk. Buckmaster may be wrong about what happened. The unsettling part is that, from the outside, neither he nor anyone else can currently prove it either way.
The acceleration nobody priced in
Step back from the dispute and look at the summer as a whole. In August, two mathematicians using AI tools settled the Euler blowup question that had resisted the field for decades. In September, a swarm of agents extended the frontier to forced Navier-Stokes in 88 hours. The debate we were having barely a year ago, whether large language models could contribute anything at all to research level mathematics, did not get resolved so much as quietly expire. A single internal model, multiplied by compute, just operated at the frontier where a few hundred human specialists work, and it got there before most of them heard the race had started.
And the model itself is a character in this story. OpenAI says the system that did this is internal only, "significantly more capable" than anything it has shipped, and was trained in the weeks immediately before the run. Whatever one thinks of the company's choices, the gap between what the public can use and what exists inside frontier labs has never been demonstrated this concretely: the public learned the model exists because of what it proved.
Answers We Can Check but Not Follow
Days before the announcement, Tao warned that an AI solution whose search process stays a black box would have a value to mathematics "close to zero", and he has since described "a very strange and unprecedented decoupling between getting answers and getting understanding". The proof is 166 pages of machine generated argument; early readers of the related AI written proofs described the prose as barely readable. Lean can certify that every step follows. No tool can certify that a human being has absorbed why.
This is the line that actually matters, more than any Millennium Prize. Mathematics has always assumed that a proof is both a certificate and an explanation. AI at this scale splits those roles apart: we can now hold correct answers we do not understand, verified by machines, produced by machines, at a pace set by machines. If the curve continues, the interesting human jobs concentrate at the two ends of the pipeline: choosing which questions are worth 10 million dollars of tokens, and doing the slow, unglamorous requirements review that decides what was actually proved. The middle, the part we called genius, is becoming infrastructure.
The Clay Institute will take its time, as it should; its two ends of the pipeline are precisely the human ones. Cordoba, whose forcing technique underpins the whole route, joked that he does not use AI because he has Martinez-Zoroa. The joke lands because it is already nostalgic. The honest summary of September 2026 is this: an AI probably did resolve two of the four official statements of a Millennium Prize problem, the hardest version of the problem still stands, no independent body has confirmed anything yet, and the most important open question raised this week is not about fluids at all. It is whether the humans in the loop, mathematicians, institutions and the rest of us, can build trust mechanisms as fast as the machines are building results.
Sources and Further Reading
The announcement, the swarm figures and OpenAI's statements are reported by the BBC. OpenAI's manuscript "Finite Time Blowup for Navier-Stokes" and the accompanying Lean certificates are public on GitHub. Tristan Buckmaster's allegations, the timeline of the dispute and OpenAI's responses are covered in detail by TechCrunch, and the mathematical context of the August Euler breakthrough by Scientific American. Quotes from Terence Tao and the Clay Institute reflect their public statements as reported in the days around the announcement.
Published September 2026. This is an opinion piece and analysis based on public reporting at the time of writing; the verification status of the proof and the dispute described here are evolving stories. CodeHelper has no commercial relationship with the companies mentioned.