OpenAI Says 10,000 Agents Found a Finite-Time Navier–Stokes Blowup

OpenAI says a system of on the order of 10,000 concurrent agents, running an unreleased internal model, produced a purported solution to the Navier–Stokes existence and smoothness problem in about 88 hours, using approximately 130 billion output tokens and 2.7 million messages on that problem alone. The 165-page PDF states a specific theorem: for every positive viscosity, a 3D incompressible flow that starts from rest can develop unbounded velocity in finite time while kinetic energy stays bounded, under a smooth, compactly supported force. That is OpenAI's claim. The Clay Mathematics Institute has not awarded the Millennium Prize, and independent mathematical acceptance is not settled.

Three records, not one. The September 8 company post is the process account. The PDF is the theorem statement. Noam Brown's September 17 interview with Dwarkesh Patel is how one OpenAI researcher talks about multi-agent compute. The interview is not peer review of the math.

What the PDF actually claims

Theorem 1.1 in Finite Time Blowup for Navier–Stokes is not “fluids explode.” It is a construction. For every viscosity ν > 0 there exist a smooth force compactly supported in space and time, a compact set in R³, and smooth velocity and pressure on [0, 1) that start from rest, stay supported in that compact set, keep kinetic energy uniformly bounded, and send the velocity's L∞ norm to infinity as t approaches 1.

OpenAI maps that construction to alternatives C and D in Charles Fefferman's official Clay writeup: breakdown on R³, and the same on the periodic torus T³. Those are the two disproof variants. A and B, the global existence-and-smoothness statements with zero force, are not what this paper claims to prove.

The force is part of the stated problem. Clay's C and D allow a smooth external force. The PDF is explicit that the fluid starts at rest and that the force is smooth. A blowup produced by feeding the equations an infinite force would be uninteresting. The claimed work is that the residual stays smooth while the velocity does not.

The process numbers, from OpenAI's post

Training of the internal model began August 28. OpenAI says the model is still training and is significantly more capable than the public GPT-6 Astra. The Navier–Stokes effort itself started September 1, after a rumor the company later tied to Levent Alpöge (Anthropic) and Tristan Buckmaster (NYU).

Agents were grouped. They could talk inside a group. Tools on the post: a cached copy of the internet, and code execution. OpenAI says it kept the same monitoring and isolation it uses on frontier evaluations.

Different groups got different problem statements. For Navier–Stokes that meant variants A and B (existence/smoothness) and C and D (breakdown). After groups explored, Codex consolidated useful intermediate results and those insights were fed back. The group that found the claimed resolution was steered that way.

OpenAI also asked the system for an “easier” cousin: blowup for the Euler equations, Navier–Stokes with the viscosity term removed. Agents produced an unforced 3D Euler regularity disproof. That run: nearly 100 agents, about 50 hours. Separate PDF.

• Navier–Stokes group: on the order of 10,000 concurrent agents.
• Resolution: Saturday, September 5, about 88 hours after the first agents launched.
• Lean formalization and verification: another 17 hours, via GPT-6 Astra, finished September 6.
• On Navier–Stokes alone: 2.7 million messages, about 130 billion output tokens.
• Across all attempted problems: 4.9 million messages, about 300 billion output tokens.
• Writeup: 165 pages, references on the last page of the PDF.
• Lean repo: openai/NavierStokesAndEuler.

A Lean certificate is a machine-checkable object. It is not the same thing as the field agreeing that the informal theorem, the formal statement, and Clay's C/D are the same claim, or that the argument is correct. That work is still for mathematicians.

Clay has not handed over the prize

OpenAI's post says the goal is to report model progress, and that the company does not intend to claim the Millennium Prize. Do not read that as Clay declining a solved problem. Read it as OpenAI declining to file.

Clay's September 11 note is careful. The institute “shares in the excitement” at an announcement that the problem has “apparently been settled.” Then it points at its own rules: evaluation and assignment of credit are deliberately unhurried. Updates later. No $1 million. No named solver. No statement that the manuscript has been refereed.

Prize rules still require a qualifying publication and a long acceptance clock. A company blog, a PDF on a CDN, and a GitHub repo are the start of scrutiny.

Buckmaster and Alpöge, as OpenAI tells it

After Lean verification on September 6, OpenAI says it contacted Buckmaster and Alpöge, offered a concurrent release, and offered to recognize their priority. In those talks the company learned their result was a resolution of forced Euler, not the forced Navier–Stokes construction OpenAI is posting. OpenAI says it offered them the prompts and, later, the proof, and that it recognizes their priority on forced Euler.

That is the primary's contact account. It is not an adjudication of priority, and it is not a claim that the two proofs are the same. Forced Euler and forced Navier–Stokes with positive viscosity are different theorems. Treat secondary writeups of the dispute as secondary.

Brown: multi-agent is parallel test-time compute, and it is not the story

On September 17, OpenAI researcher Noam Brown sat with Dwarkesh Patel. Patel opened with the company's numbers: 10,000 agents, 130 billion tokens, 88 hours. Brown's framing is compute, not Clay.

Serial test-time compute has a latency wall. Multi-agent, in his account, is how you spend that compute in parallel. Less efficient than one agent holding the whole context. Faster wall-clock if the domain parallelizes. He says math parallelizes reasonably well. A novel would not.

“I wouldn't even attribute 10% of the credit to multi-agent. The reality is that OpenAI has trained a very powerful model.”

Noam Brown, Dwarkesh Patel interview, September 17, 2026

Published Ultra Mode plots, from the GPT-5.6 launch, go to 1, 4, and 16 agents. Default Ultra is four. Some benchmarks: four agents finish about twice as fast at about 2× token cost. Sixteen continues the pattern, a bit less efficient. Brown is explicit that they do not have that science at 10,000. One weekend run is one data point. They have not timed a single agent on Navier–Stokes. A thorough ablation at that scale is, in his words, too expensive. Do not turn 10,000 into a measured speedup.

He also refuses the clean story that 10,000 agents coordinated like 10,000 people. “We think it helped. We don’t actually have good measurements.” It is possible, he says, that 10,000 humans still coordinate better.

The public multi-agent example he reaches for is “unfortunately the Hugging Face one.” That is the July 2026 ExploitGym incident Ultrathink already covered: about 700 agents, 70,000-plus messages, Hugging Face compromised. Brown is talking organization that emerged from loose tools, not endorsing that run.

What the interview is actually about

Patel wants a read on recursive self-improvement. The math timeline he sketches: contest problems in 2024, IMO gold in 2025, open problems, then a Millennium claim. Brown says the 10×-per-year jump in “how long would a human mathematician need” made a 2026 Millennium result earlier than he expected. He had been looking at 2028.

He also pushes back on “the models replaced mathematicians.” Jagged: strong at well-scoped, checkable problems; weaker at posing new ones and deciding which branches of mathematics are worth opening. Treat that as Brown's opinion about models, not a referee report on Theorem 1.1.

RSI, in his telling, looks more like math than like Go in one respect: the objective is measurable. It looks less like math in another: you still have to run experiments, and compute remains a bottleneck. He does not claim Navier–Stokes proves an intelligence explosion. Patel is the one pumping the intuition. Brown calls the compute picture “pretty accurate” and then puts experiments back in the way.

What this is not

• Not a Clay award. Not “the Millennium Prize was won.”
• Not a proof of Clay A or B (global smooth solutions for every smooth unforced datum).
• Not an unforced Navier–Stokes blowup. The NS paper uses a smooth force.
• Not a statement about physical water or air. Infinite velocity is a feature of the continuum model under the constructed data, not a weather forecast.
• Not a measured 10,000-agent ablation. Brown said they do not have that science.
• Not Brown peer-reviewing the PDF. He is talking about agents and credit.
• Not a finding that OpenAI used Buckmaster and Alpöge's unpublished proof. The primary says the company contacted them after its own Lean pass and then learned their result was forced Euler.

The live objects to check are the PDF, the Lean repo, Fefferman's C/D statement, and whatever the field publishes next. Until Clay finishes its unhurried process, the accurate sentence is the short one: OpenAI posted a claimed C/D construction, produced by a very large agent run on an unreleased model. The prize is still on the table.