OpenAI Navier-Stokes Claim: 10,000 Agents, Proof and Dispute

A smooth force drives a bounded-energy flow to infinite speed

OpenAI's paper constructs, for every positive viscosity, a solution of the three-dimensional incompressible Navier-Stokes equations that starts from zero velocity and is driven by an external force that is smooth and compactly supported in space and time. The velocity and pressure remain smooth before time 1 and the flow keeps uniformly bounded kinetic energy, but its maximum velocity becomes unbounded as time approaches 1. The paper says this establishes the breakdown alternatives C and D in Charles Fefferman's official Millennium Prize problem statement, for both all of three-dimensional space and a periodic three-dimensional torus. [5][6][1]

The construction concentrates a rotating vortex into an ever smaller core. Oscillatory pulses around that core create nonlinear momentum fluxes that cancel the singular part of the force the background flow would otherwise require, leaving an external force that stays smooth even as the velocity diverges. It is an existence proof for a deliberately constructed idealized flow, not a simulation showing that ordinary water, air, or every Navier-Stokes solution develops infinite velocity. [5][1]

Ten thousand agents searched for 88 hours

OpenAI says it began the effort on September 1 after hearing a rumor that other researchers had resolved two Millennium Prize problems. It set groups of agents on different approaches, then used Codex to consolidate useful intermediate results across groups. Roughly 10,000 agents reached the Navier-Stokes result on September 5, 88 hours after the first agents started. OpenAI reports 2.7 million messages and about 130 billion output tokens for the Navier-Stokes work, followed by another 17 hours in which GPT-6 Astra formalized and verified the proof in Lean. [2][1]

The result was not one model answering one prompt. It was a managed search over many candidate arguments, with humans choosing how groups were organized and when their work was combined. The system that found the argument was itself unreleased and, according to OpenAI, significantly more capable than GPT-6 Astra. The public artifacts make this result inspectable, but they do not make the underlying search system reproducible outside OpenAI. [2][6]

Machine-checked is not yet Clay-accepted

OpenAI published the 166-page paper and a Lean repository that formalizes both breakdown alternatives and provides build instructions plus a separate comparator-checking path. Lean checks that the encoded theorem follows from the encoded definitions and prior results. Quanta notes the remaining human obligation: establish that the formal statement is logically equivalent to the mathematical problem people meant to solve. [5][6][1]

Fefferman told EL PAIS that the result does answer the problem as he formulated it, while also saying he wanted a conventional prose rewrite and close reading by several experts before being fully convinced. Clay still listed Navier-Stokes as active when checked on September 12. Its rules require publication in a qualifying outlet, at least two years of rigorous examination, general acceptance in the mathematics community, and then Clay's own review. OpenAI says it does not intend to claim the $1 million prize. [11][7][8][2]

The theorem and the credit dispute are separate questions

Tristan Buckmaster and Levent Alpoge announced related AI-assisted results on smooth-forced blowup for incompressible porous media, Boussinesq, and three-dimensional Euler. Buckmaster credited Diego Cordoba and Luis Martinez-Zoroa with the basic cascade program and said his team had used Claude and Codex. His statement did not claim a full Navier-Stokes solution: it described a hypo-dissipative Navier-Stokes result whose Lean check and presentable paper were not finished. Quanta reports that both teams built on the Cordoba-Martinez-Zoroa line of work. [9][1]

Buckmaster said OpenAI began after word of his team's progress had spread, arrived at the same uncommon smooth-forcing route, and discussed a paper arrangement that would omit Alpoge, followed by a remark he understood as a career threat. He also wrote that he had not seen OpenAI's proof, did not know whether his team's Codex data was used, and was not making an accusation about that. Sebastien Bubeck denied trying to remove Alpoge from credit for Alpoge's own work, said the disputed authorship concerned a rewrite of OpenAI's proof, and apologized for the career remark. Those private conversations have not been independently adjudicated. [9][10]

OpenAI's September 10 update says an internal investigation found that Buckmaster's Codex prompts could not have influenced the system, including through training, and that no one involved saw the outside work before its public release. That is OpenAI's account of its own systems, not a public technical audit. The mathematical claim can survive or fail on its proof while questions about priority, attribution, and the use of confidential research follow a different evidentiary path. [2][10]