Ten thousand AI agents, 88 hours, and a fight over who got there first

Eighty-eight hours, and roughly 10,000 AI agents running in parallel. That is what OpenAI says it took to produce a proof that a smooth fluid at rest under the Navier-Stokes equations can develop a singularity in finite time. The proof was formalised in Lean and published.

Navier-Stokes existence and smoothness is one of the seven Clay Millennium Prize problems, and each carries a 1 million dollar reward. OpenAI added the caveat itself before anyone else could: it does not intend to claim the money.

What the OpenAI Navier-Stokes proof actually claims

The model behind the run is unreleased and internal. Its agents coordinated with each other rather than working in isolation, which is what separates this from one very long reasoning session. Formalising the result in Lean matters too, because a Lean proof can be machine-checked instead of taken on trust.

Running 10,000 agents for 88 hours is an inference bill rather than a training one, which is its own kind of expensive and getting more so.

The Clay Mathematics Institute has certified nothing. The problem is still listed as open. Clay rules require publication in a recognised journal, independent verification by the mathematical community, and a two year waiting period before a submission counts at all. The million dollars was never on the table this week.

Then the credit fight started

NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpoge had spent close to a year on related fluid equations, using an approach Buckmaster describes as unusual. He questions whether OpenAI arrived at the same method independently, and he notes that the OpenAI effort began on 1 September, after word of their progress had circulated.

OpenAI denies accessing private prompts or unpublished work. The published page now credits both men for concurrent work on the forced Euler problem, which reads less like a correction and more like a settlement.

OpenAI researcher Sebastien Bubeck said the initiative started because rumours suggested Anthropic had resolved two Millennium problems. Which tells you the motivation was competitive before it was mathematical.

Nobody has a rule for this

The maths will get checked. Lean makes that tractable, and the community will either confirm the result or find the hole in it. That part is boring in the good sense.

The attribution is the part that matters. A researcher spends a year on an unusual approach and discusses it the way researchers normally do. A lab reads the same public literature and lands somewhere similar in under four days. No rule was obviously broken, and no rule obviously exists.

Academic priority has always been settled by publication date, and publication date assumes both parties move at roughly human speed. Ten thousand agents for 88 hours breaks that assumption without offering a replacement. We saw a smaller version of the same compression when four frontier models shipped inside 72 hours, and the release cadence this month has not slowed since.

What to watch

Two things decide how this gets remembered. Whether the Lean proof survives community verification over the coming months, and whether Clay updates its guidance on machine-assisted submissions, given that its rules predate agent swarms entirely. Buckmaster and Alpoge publishing their own version would give the record something concrete to compare against.

The prize will not be claimed. The argument about who got there first will outlast it.