On September 8, 2026, OpenAI published what it describes as a solution to the Navier-Stokes existence and smoothness Millennium Prize Problem. An internal model, said to be significantly more capable than GPT-6 Astra, coordinated roughly 10,000 agents to construct an analytical proof. GPT-6 Astra then helped formalize and verify it in Lean.
If expert review confirms the argument, the result would settle a question that has resisted mathematicians for roughly 90 years. That conditional matters. A machine-checked Lean proof is strong evidence that the formal statements follow from their encoded assumptions, but it does not remove the need for mathematicians to inspect whether those assumptions and definitions match the original problem.
OpenAIWhat the OpenAI Navier-Stokes proof claims
The Navier-Stokes equations describe fluid motion. The Millennium problem asks whether smooth three-dimensional incompressible flow must stay smooth, or whether velocity can become unbounded in finite time. OpenAI's paper argues for the second outcome under the forced formulation: a smooth fluid beginning at rest develops a finite-time singularity while its energy remains finite and the external force remains smooth.
The proposed mechanism is a vortex that spirals inward and stretches along its axis. Rotation accelerates as the central region shrinks, while pressure, viscosity, acceleration, and momentum-transfer terms cancel precisely enough to keep the applied force smooth. OpenAI says this establishes statements C and D in the Clay Mathematics Institute's official formulation.
The reported proof effort took 105 hours across two phases
Elapsed time reported by OpenAI. The 88-hour agent search and 17-hour Lean formalization were sequential phases, not independent benchmark runs.
Source: OpenAI, September 8, 2026. Method: direct visualization of vendor-reported elapsed hours.
How a 10,000-agent research system worked
OpenAI says it began the effort on September 1 after hearing rumors of progress on two Millennium problems. Agent groups received different formulations and could read a cached internet, run code, communicate within groups, and exchange promising intermediate ideas. Nearly 100 agents first produced a related result for the unforced Euler equations. OpenAI then redirected more compute toward Navier-Stokes and used Codex to consolidate useful ideas across groups.
The successful agent phase ended on September 5 after about 88 hours. Lean formalization and verification took another 17 hours. Across Navier-Stokes alone, OpenAI reports 2.7 million messages and about 130 billion output tokens. These are process measurements from the developer, not an independent efficiency benchmark, but they show the mechanism clearly: unusually broad parallel search followed by formal checking.
Formal verification is not final acceptance
Lean can verify every inference inside a formal proof once the mathematics has been translated into the proof assistant. That makes hidden algebraic slips much harder to overlook. It does not automatically establish that the formal theorem is equivalent to the Clay problem, that every imported result applies as intended, or that the broader mathematical community accepts the construction.
OpenAI therefore published three inspectable artifacts: a prose writeup, a paper, and the Lean repository. The appropriate status is "proposed solution," not an already awarded Millennium Prize. OpenAI also says it does not intend to claim the prize. Clay's normal recognition rules require publication, a waiting period, and general acceptance by the mathematical community.
The concurrent-work and data-trust issue
The research timeline also produced a dispute about credit. OpenAI says the rumor that prompted its search concerned work by NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge on the related forced Euler problem. The lab says its agents did not access their specific work before public release, while acknowledging it cannot completely rule out an indirect contribution from de-identified product data used to improve its models. Axios reported objections around how OpenAI handled the concurrent work and proposed attribution.
Those claims require separate evaluation from the proof itself. A valid proof would remain mathematically important, while the episode still raises a governance question for AI-assisted science: researchers need clear assurance about how confidential prompts, unpublished ideas, and model-improvement data are separated from a lab's own research programs.
What this result changes
The most defensible conclusion today is not that AI has officially solved Navier-Stokes. It is that a frontier lab has released a detailed, formally checked candidate solution produced through massive agent coordination, making the evidence open to expert scrutiny. That is a sharper research milestone than a benchmark score because specialists can inspect the actual artifact.
The result also extends the story behind GPT-6 Astra. Capability is moving from answering known questions toward searching research spaces, coordinating thousands of attempts, and generating claims that can be formally tested. The next decisive evidence will come from mathematicians reviewing the equivalence, novelty, and correctness of the proof.
Sources:
- On the Navier-Stokes Millennium Prize Problem | OpenAI
- OpenAI paper on the Navier-Stokes solution
- Lean formalization repository | GitHub
- Official Navier-Stokes problem formulation | Clay Mathematics Institute
- Rules for the Millennium Prize Problems | Clay Mathematics Institute
- OpenAI's math result and the credit controversy | Axios
Image credits
Header and in-body photograph: 1515 Third Street in San Francisco, used as OpenAI's headquarters, photographed by Coolcaesar via Wikimedia Commons, licensed under CC BY 4.0. Reviewed September 9, 2026. The photograph does not depict the proof effort.