There are seven Millennium Problems — mathematical questions so difficult that the Clay Mathematics Institute offered a million-dollar prize for each one in the year 2000. In the 26 years since, only one had been solved. A Russian mathematician named Grigori Perelman proved the Poincaré Conjecture in 2003, then declined the prize and the Fields Medal that came with it, and more or less vanished from public life. For close to thirty years, these six remaining problems have persisted, eluding the most gifted mathematical thinkers on Earth.
On September 8, 2026, OpenAI published a claimed solution to a second one.
The Navier-Stokes existence and smoothness problem concerns the equations that govern how fluids and gases move — how water flows around a stone, how air moves over a wing, how smoke rises in a column before it eddies. The equations themselves are not disputed. They are among the most successful in physics. It has never been established if they consistently stay mathematically well-behaved, or if there are fluid motion configurations where the equations fail — causing the mathematics to yield an infinite value within a finite time, known as blowup.
OpenAI’s proof says blowup happens. Under smooth initial conditions and smooth external forces, the Navier-Stokes equations generate a vortex that contracts and spins increasingly fast — a singularity where the velocity becomes locally infinite, yet the fluid’s total energy remains finite. The proof is 165 pages long, formalized in Lean 4, and deposited on GitHub. Anyone can verify it without trusting OpenAI at all.
How It Was Done
OpenAI initiated the project on September 1st after rumours emerged linking Tristan Buckmaster, a mathematics professor at New York University, and Levent Alpöge, a mathematician at Anthropic, both of whom had been independently addressing similar problems. More on that in a moment.
“The agents arrived at their resolution on Saturday, September 5, about 88 hours after the first agents were launched. Lean formalization and verification took an additional 17 hours via GPT-6 Astra.”
The model powering the agents is not GPT-6 Astra. OpenAI describes it as an internal system that is far more capable than Astra, an unreleased model without a public name. Independent coverage has referred to it informally as Astra-next. It orchestrated roughly 10,000 coordinating agents, each capable of reading from a cached version of the internet, running code, and exchanging messages with the other agents.
“OpenAI encouraged different groups of agents to explore a diversity of approaches. After some time, they cross-pollinated the agent groups by using Codex to consolidate the most useful insights from each agent group. These follow-up prompts drew on the agents’ own intermediate results. The group that found the solution to Navier-Stokes was guided in such a way.”
The overall computational expense for the Navier-Stokes problem amounts to roughly 130 billion output tokens and 2.7 million agent messages. “Independent analysts at Latent Space put the cost at more than $40 million in compute” — making it, by a wide margin, the most expensive single AI research run ever publicly disclosed.
OpenAI said it does not intend to claim the $1 million prize.
What the Proof Actually Shows
The specific mathematical result is worth understanding in plain terms because the way it’s been described in some coverage has created confusion about what was actually proved.
The Navier-Stokes Millennium Problem has two resolutions. One would show that the equations always remain smooth and well-behaved — that no matter what initial conditions you choose, the fluid motion they describe never breaks down. The other would demonstrate that certain configurations exist in which the equations fail — where a singularity emerges.
OpenAI proved the second case. The mathematical object they constructed is a 3D fluid — not a real fluid you could pour in a glass, but a fluid in the mathematical sense defined by the problem — with a specific initial configuration under smooth external forces.
“OpenAI’s model found it can, describing a configuration in which a vortex tightens and spins ever faster — a phenomenon mathematicians refer to as finite-time blowup — while the fluid’s energy stays bounded throughout.”
This solves the Clay problem. But it is important to be precise: the problem had two answers, and the one OpenAI found — blowup exists — is in some sense the more troubling one for physics. It suggests that in certain extreme mathematical configurations, the equations governing real fluid motion might produce values that don’t correspond to anything physically observable. The practical consequence for engineering and physics is likely limited. The mathematical consequence is profound.
The Lean formalization is the component that elevates this beyond any benchmark result. Lean is a formal proof assistant whose verification kernel is a compiler. “Mathematics has something almost no profession has: a cheap, mechanical way of checking whether an answer is right.” Either the proof is logically complete, and the Lean checker accepts it, or it isn’t. The sorry count in OpenAI’s formalization: zero. Every logical step, verified.
The Controversy That Started Before the Proof Was Published
The cleaner version of this story would have OpenAI solving a Millennium Problem in 88 hours and the mathematical world responding with uniform amazement. That is not what happened.
“Before Tuesday’s announcement, competing researchers challenged OpenAI’s account of how the work unfolded.”
Tristan Buckmaster and Levent Alpöge had been working independently on the forced Navier-Stokes problem. Alpöge is employed at Anthropic. Their method tackled the forced Euler equations — a connected yet separate issue regarded as a preliminary step toward the complete Navier-Stokes solution. According to Buckmaster, Alpöge received information suggesting that details of their unpublished progress had made their way to OpenAI.
“Buckmaster explained that the tech giant’s route to the solution was similar to their own work, which is “not the direction one arrives at in a few days by giving a model the problem statement.””
OpenAI’s response was detailed and direct. “The company stated that they did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed. OpenAI offered them visibility into all of the prompts used and later to see the proof. They recognized the priority of Buckmaster and Alpöge’s work on forced Euler and congratulated them on their remarkable mathematical achievement.”
What Alpöge and Buckmaster proved using Anthropic’s model — finite-time blowup for the forced Euler equations — is a significant independent result. The dispute isn’t about who is correct in the mathematics. Both proofs appear to be valid. It is about the provenance of the approach and whether a company that trains on internet-accessible data can credibly claim independence from a researcher’s unpublished methodology when the methodology points in an almost identical direction.
That question has no clean answer. This offers a glimpse of the intellectual property challenges that will shape AI-assisted research in the coming years.
What Terence Tao Said — And What He Published the Day Before
The day before OpenAI’s announcement, on September 7, Terence Tao published a paper on finite-time blowup results for three related equations: the incompressible porous medium equation, the Boussinesq equations, and the incompressible Euler equations. Tao had not been working with OpenAI. He worked alone, employing his personal methods, and released his work independently.
His work and OpenAI’s work — both appearing within 24 hours — are not the same result. Tao’s paper addresses Euler, not Navier-Stokes. But the convergence of three independent groups arriving at blowup results for related equations within days of each other suggests the field was ripe for this discovery, that multiple paths existed, and that the specific direction of OpenAI’s approach may have been discoverable through legitimate means with no data access controversy.
Tao’s additional input to this story is cultural rather than mathematical. He has said publicly that the question he worries about is not whether AI can prove theorems — he accepts it can and will. The question he worries about is whether indiscriminate AI-assisted proof production will transform mathematics into what he called a production quota game: where the goal becomes generating Lean-verified statements rather than developing genuine mathematical understanding. He is not saying the Navier-Stokes result is wrong. He argues that volume and validity differ, and that the math community must establish new standards for meaningful insight beyond merely verifiable results.
That is a convincing argument, and you don’t have to question the evidence to consider it seriously.
The Cost Is the Signal
The $40 million compute bill is easy to dismiss as corporate excess. It is worth reading as a unique signal.
Two weeks ago, in the article about Astra’s ten open mathematical problems, I noted the total compute cost for those results was approximately $2,000 at current API rates. A century’s worth of mathematical work, including a problem open since 1999, for $2,000.
The Navier-Stokes result cost $40 million. That costs 20,000 times more. This reflects something real: the problem is qualitatively harder, requiring the coordination of ten thousand independent agents over nearly four days, each pursuing different strategies, their intermediate results being synthesized and redirected by Codex in real time.
But the trajectory to watch is not the absolute cost. It is the direction the cost is moving. “The labs will not be running ten thousand agents in 2028. They will run a million, on something correspondingly harder.” Unit prices for inference have been declining faster than any other cost curve in the technology industry. The $40 million run of September 2026 becomes the $4 million run of 2027 and the $400,000 run of 2028 — not because the problems get easier, but because the compute gets cheaper and the agents get smarter. The limits of what AI can handle mathematically are not set in stone. It is moving upward.
Sam Altman’s tweet after the result was published nominated the next target: room-temperature superconductors. The challenge of achieving superconductivity at ambient conditions has been one of the defining unsolved problems in condensed matter physics for decades, with false claims and retracted papers embarrassing the field repeatedly. It is a different category of problem from pure mathematics — empirical rather than formal, requiring physical experiments that cannot be fully replaced by computation. However, the goal of the ambition is well-defined.
Is It Actually Solved?
The Clay Mathematics Institute has not yet commented officially. This is appropriate. The Institute takes its time, requiring a minimum of two years for the community to review and approve the proof before awarding a Millennium Prize. Perelman’s proof of the Poincaré Conjecture was verified through multiple independent confirmations over several years before the Institute considered the matter settled.
The Lean formalization changes the shape of that process, not the necessity. What Lean confirms is logical validity — that the proof is internally consistent and every step follows from the previous one. What it cannot automatically verify is whether the problem, as formalized, exactly matches the Clay problem as initially presented. The mapping between the informal Clay statement and the formal Lean encoding is itself a mathematical question, and one that requires human mathematical judgment to verify. OpenAI has been explicit about this.
What we have right now: a 165-page proof, a verified Lean formalization, and the judgment of individual mathematicians who have reviewed it that the approach appears sound. The Quanta Magazine coverage included responses from Charles Fefferman — who wrote the Clay problem’s official description — and Tristan Buckmaster, both engaging seriously rather than dismissing it.
That is not the same as a confirmed solution. It is close enough to be historically significant.
The Remaining Five
Before September 8, 2026, two of the seven Millennium Problems had been resolved: the Poincaré Conjecture by Perelman in 2003, and now Navier-Stokes existence and smoothness problem by a system of AI agents in 88 hours. The five that remain are: the Birch and Swinnerton-Dyer Conjecture, the Hodge Conjecture, P versus NP, the Riemann Hypothesis, and Yang-Mills existence and mass gap.
The P versus NP problem has the most direct consequences for the world beyond mathematics. It asks whether problems that are easy to verify are also easy to solve. If P equals NP — if the answer is yes — then cryptography as currently practised is broken. Every encrypted message, every secure financial transaction, every private communication would be theoretically vulnerable. If P does not equal NP — the answer most mathematicians expect — then the current foundations of digital security are confirmed.
It is the one problem on the list that the world probably hopes AI does not solve quickly.
What do you think the Lean verification changes about how you read this result? Does machine-checked proof satisfy you, or does it feel like something important — the human reconstruction of why something is true — is still missing? I’m uncertain how to answer that question myself.
If you read this far and want every piece from the Nov Tech archive in the order it was written, follow us and subscribe.