10,000 AI Agents Wrote a Proof for a Millennium Problem

OpenAI has released a 166-page argument claiming that a smooth, three-dimensional fluid can develop an infinite-speed singularity in finite time. If mathematicians accept the proof, it would settle the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. The result was generated by a large AI-agent system and separately encoded in Lean, but formal checking is not the same as final acceptance by the mathematical community.

KEY NUMBER
OpenAI reports that the Navier-Stokes effort used roughly 10,000 concurrent agents, 2.7 million messages and about 130 billion output tokens.

What the proof is trying to establish

The September 8 announcement argues for the counterexample route in the Clay formulation. The constructed fluid begins at rest, receives a smooth external force and keeps finite total energy, yet a tightening vortex accelerates without bound in finite time. Such a singularity would show that smooth solutions need not remain smooth forever.

This is a statement about the equations, not a prediction that real water can reach infinite speed. A singularity marks the point where the continuum model breaks down and a different physical description would be needed.

How the AI search was conducted

OpenAI says the agent group worked for about 88 hours, then GPT-6 Astra spent another 17 hours translating and checking the result in the Lean proof assistant. The company has published the analytical paper and Lean repository.

The scale is extraordinary, but it also makes the experiment difficult to reproduce without similar computing resources. It extends an earlier NewTqnia report about AI contributing a new construction to an Erdős problem from useful mathematical assistance to a claimed resolution of a famous open theorem.

Formal verification is not the final verdict

Lean confirms that each encoded step follows from the encoded assumptions. It does not by itself establish that the formal statement perfectly captures the intended Millennium problem or that every imported assumption is acceptable. Mathematicians interviewed by Quanta therefore describe the result as potentially historic while continuing a line-by-line human review.

The Clay Mathematics Institute still lists the problem as unsolved, and OpenAI says it will not claim the prize. A formal community verdict may take time.

A credit and data question remains open

OpenAI began the project after hearing rumors of related work by mathematicians Levent Alpöge and Tristan Buckmaster. Buckmaster says their unpublished work used Codex and asks whether de-identified product data could have influenced later models. OpenAI says no specific user data was accessed, while acknowledging that it cannot completely rule out indirect influence from de-identified usage data. That is not evidence of copying, but it leaves provenance and credit questions that deserve independent scrutiny.