OpenAI's Navier-Stokes Proof Explained (September 2026)
The Short Answer
On September 8, 2026, OpenAI announced that an unreleased internal model — running as a swarm of roughly 10,000 coordinating agents — produced a proof that the three-dimensional Navier-Stokes equations can develop a singularity in finite time. If it holds, it resolves the Navier-Stokes existence and smoothness problem, one of the seven Clay Millennium Prize Problems, by disproving global smoothness under the “forced” formulation (statements C and D of the official problem).
Three things are true at once, and the story only makes sense if you keep them separate:
| Claim | Status as of September 9, 2026 |
|---|---|
| A proof exists and is public | Yes — 166-page manuscript + Lean repository released September 8 |
| The proof has been independently verified | No — repo review status is “self-assessed”; no outside reproduction yet |
| OpenAI will claim the $1M prize | No — OpenAI says it does not intend to |
| The result is undisputed | No — credit and provenance dispute with Buckmaster and Alpöge |
What Was Actually Proved
The Clay formulation has four options. A and B ask you to prove smooth solutions always exist (no external force, in free space or a periodic box). C and D ask you to exhibit a breakdown — smooth initial data, a smooth external force satisfying decay conditions, and a solution whose velocity becomes unbounded in finite time.
OpenAI’s Theorem 1.1 targets C and D. For every positive viscosity, it constructs a fluid that starts at rest, is driven by a smooth force compactly supported in space and time, and reaches unbounded velocity at a finite time while kinetic energy stays bounded. Corollary 10.6 gives the periodic version. It does not claim unforced blowup — that distinction matters for anyone summarising this as “Navier-Stokes is solved.”
The mechanism is a contracting vortex: a spinning core that spirals inward and stretches like spaghetti, shrinking in width faster than length. The hard part — the reason this stayed open for roughly 90 years — is making the acceleration, pressure, momentum-transfer and viscosity terms all blow up yet cancel precisely enough that the external force stays smooth.
How the Agent Swarm Worked
OpenAI’s own account (published September 8) is unusually detailed, and the numbers are worth quoting exactly because they are company-reported, not audited:
- Model: an internal model in training since August 28, described as “significantly more capable than GPT-6 Astra.” It is not released and has no public name.
- Start: Tuesday, September 1, after OpenAI heard rumours that two Millennium Prize problems had been resolved.
- Structure: agents subdivided into groups that could communicate internally, with tools to read a cached copy of the internet and run code. Different groups got different variants of the problem (A/B for a proof, C/D for a disproof).
- Warm-up: nearly 100 agents resolved the unforced Euler regularity problem (Navier-Stokes without viscosity) in about 50 hours. That result was then fed to the Navier-Stokes groups.
- Cross-pollination: Codex was used to consolidate the most useful insights across groups; when a further-trained version of the model became available mid-run, agents were switched to it.
- Result: the winning group was on the order of 10,000 concurrent agents, reaching the resolution on Saturday, September 5 — about 88 hours after launch.
- Formalization: 17 additional hours of Lean work using GPT-6 Astra.
- Scale: across all problems attempted, 4.9 million messages and ~300 billion output tokens; for Navier-Stokes specifically, 2.7 million messages and ~130 billion output tokens. Chief research officer Mark Chen put the compute cost in the “millions of dollars.”
Sébastien Bubeck has said the human team lacked research-level fluid-dynamics expertise and could not meaningfully contribute mathematical content — the orchestration was human, the mathematics was the model’s.
Why It Is Disputed
Verification. OpenAI described the result on a press call before publishing. By the evening of September 8 the manuscript and Lean project were public, which changed the evidence from “trust us” to “inspect this.” But the repository’s formalization.yaml labels review status as self-assessed. Lean can only certify the statement it encodes; experts still need to confirm the formal theorem matches the paper, the build completes without sorry, and only the standard three axioms are used. None of that had been reproduced independently as of September 9.
Credit. Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) had posted preprints — with Lean formalizations — proving finite-time blowup for the incompressible porous medium equation, 2D Boussinesq, and forced 3D Euler. Buckmaster alleges OpenAI learned of their direction, pursued it, and in September 6 calls pressed him over publication and authorship, including excluding Alpöge because he works at Anthropic. OpenAI denies this: it says neither its researchers nor its agents saw the pair’s work before public release, that no user data was accessed, and that its Euler proof is for the unforced case — a different theorem. Bubeck has apologised for his part in the exchange. These remain contested accounts of private conversations.
Prize. Clay rules require publication in a qualifying outlet, two years of scrutiny, and general acceptance. Clay still listed Navier-Stokes as unsolved on September 9. OpenAI says it will not claim the prize.
What Would Settle It
An outside group reproducing the Lean build and confirming the formal statement matches Theorem 1.1. That is a matter of weeks, not years — and it is the only update that actually changes the mathematical status. Everything else is provenance.