OpenAI Solves Navier–Stokes in 88 Hours With 10,000 Agents

On September 8, 2026, OpenAI published a research paper and accompanying Lean formalization claiming a resolution of the Navier–Stokes existence and smoothness problem — one of seven Millennium Prize Problems designated by the Clay Mathematics Institute in 2000. The result, produced by an internal model described as significantly more capable than GPT-6 Astra, establishes that a smooth, initially quiescent fluid under a smooth applied force can develop a velocity singularity in finite time, even with viscosity present. That resolves statements C and D in the official Millennium Prize formulation, answering a question that had remained open since Jean Leray's 1934 generalized-existence proof.

For practitioners tracking frontier AI capability verification, the claim matters less as a mathematical trophy and more as an operational signal: a multi-agent system running an unreleased internal model autonomously navigated one of the hardest open problems in analysis, produced a machine-checkable formal proof, and did so in under four days of wall-clock time.

The Proof and What It Claims

The solution describes a vortex that spirals inward while undergoing axial stretching — a geometry in which the central region contracts and accelerates while total energy remains finite throughout. The mathematical difficulty is that the Navier–Stokes terms — acceleration, pressure gradients, momentum transfer, and viscosity — must simultaneously diverge yet cancel with enough precision that the external forcing field remains smooth. The singularity arises from the fluid's own dynamics, not from an artificially singular forcing term introduced by hand.

Lean formalization and verification took an additional 17 hours and was performed via GPT-6 Astra after the analytic proof was completed on September 5.

The Multi-Agent System

OpenAI began training the internal model on August 28. After hearing rumors on September 1 about external Millennium Prize resolutions, the team launched a multi-agent evaluation sweep across all open Millennium Prize problems. Agents had access to a cached internet snapshot and code execution. The Navier–Stokes group ran on the order of 10,000 concurrent agents. Cross-pollination between groups was handled using Codex to consolidate intermediate insights and re-prompt downstream agents with the most productive results.

The agents first resolved a related but distinct problem — the regularity question for the Euler equations (Navier–Stokes with the viscosity term removed), in the unforced variant — using nearly 100 agents over approximately 50 hours. That Euler result served as a stepping stone: resources were shifted away from other Millennium Problems, and agents working on Navier–Stokes were explicitly prompted with the Euler resolution. The full Navier–Stokes solution arrived on September 5, roughly 88 hours after the first agents were launched.

Total resource consumption across all attempted problems: 4.9 million messages and approximately 300 billion output tokens. For the Navier–Stokes problem alone: 2.7 million messages and approximately 130 billion output tokens.

Concurrent Work and Priority

OpenAI's rumors pointed toward Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a mathematics professor at NYU. After completing Lean verification on September 6, OpenAI reached out to offer a joint announcement, at which point it discovered that Alpöge and Buckmaster had resolved the forced Euler problem — a distinct result from OpenAI's unforced Euler regularity disproof. OpenAI acknowledges their priority on forced Euler and states the two proofs differ significantly even where the mathematical territory overlaps.

OpenAI explicitly states it does not intend to claim the Millennium Prize for this result.

What the Token Budget Indicates

The 130-billion output-token expenditure on a single problem marks this as a compute-intensive research process — closer in character to pipeline-driven AI research than to single-pass generation. The internal model powering it sits above GPT-6 Astra in capability, and its training was still ongoing and improving during the September 1–5 window. That detail — a model improving mid-deployment across a live research effort — is the sharpest signal here. Whether or not the mathematical community fully ratifies this proof on first review, an agentic system producing a Lean-checkable candidate resolution to a 90-year-old problem in under four days materially resets expectations for what AI research benchmarks can credibly target.