OpenAI's 10,000 Agents Claim a Fluid-Dynamics Proof

On September 8, OpenAI announced that one of its unreleased models, running for about 88 hours across as many as 10,000 parallel agents, had produced a proof of finite-time blowup for the forced three-dimensional Navier-Stokes equations. The company published a roughly 165-page written proof alongside a Lean 4 formalization that other mathematicians can download, build and check line by line. That last part matters. A machine-verified proof is not a press release you have to take on faith.

It is worth being precise about what was and was not settled, because the headlines have blurred it. The Navier-Stokes equations describe how fluids and gases move, and one of the great open questions is whether a smooth flow can develop a genuine singularity, a point where the mathematics breaks down in finite time. The Clay Mathematics Institute attached a million-dollar Millennium Prize to the standard, unforced version of that question. What OpenAI's system tackled was the forced version, where an external term is added to drive the equations. As Quartz and others noted, that is a real and difficult result, but it is not the prize problem. Nobody is collecting a cheque.

A genuine result, wrapped in a dispute

The mathematics community's early read is that the proof looks sound, which is remarkable on its own. What is less settled is where the idea came from. By OpenAI's own account, the effort began on September 1 after the company caught wind of a rumor that a similar result was in the air. That rumor traced back to Levent Alpoge, a researcher at Anthropic, and Tristan Buckmaster, a mathematician at New York University, whose related work concerned the forced Euler equations rather than Navier-Stokes.

Buckmaster has since disputed OpenAI's framing. According to his account, reported by Forkast, OpenAI was already aware of his parallel work, followed the same narrow line of attack, and then pressed him to drop his Anthropic-affiliated collaborator from any joint credit. OpenAI's Sebastien Bubeck has said the company cannot fully rule out that the model benefited from a researcher's private notes that had passed through its tools. That is an uncomfortable admission to have to make about a proof you are presenting as a landmark.

Why the swarm is the story

Strip away the credit fight and something new remains. This was not a single model producing a clever answer. It was thousands of agents working in parallel for the better part of four days, proposing lines of argument, checking each other, and grinding a research-grade problem down to a formally verified conclusion. It is the same shift toward automated, long-horizon research that we saw when OpenAI described its agents as research interns, and when Claude formalized Fermat's Last Theorem in Lean.

None of this means human mathematicians are finished, a fear we looked at squarely last month. The interesting questions here were still framed by people, and the provenance row is itself a very human argument about who thought of what first. But the method is starting to look less like a party trick and more like a tool. When the tool can hand you a proof a computer will vouch for, the argument moves from whether to trust it to who deserves the citation. That is new ground, and the field is going to be arguing over it for a while.