← Back to feed

OpenAI announced on September 8 that an internal agentic system has produced a proof resolving the Navier-Stokes existence and smoothness problem, one of the seven Clay Mathematics Institute Millennium Prize Problems. The result shows that smooth three-dimensional fluid motion can break down — that singularities exist — answering a question left open for roughly 90 years.

The proof was accompanied by a complete Lean formalization, verified by GPT-6 Astra over 17 hours.

The Result

The Navier-Stokes equations describe how fluids move, treating a fluid as a continuous medium rather than tracking individual molecules. Since Jean Leray proved generalized solutions exist in 1934, the central open question had been whether those solutions always remain smooth or can blow up to infinite velocity in finite time.

OpenAI’s system established that blow-up can occur. The mechanism is a vortex that spirals inward and elongates axially, like spaghetti. As the central region contracts, the fluid velocity grows without bound while total energy remains finite — satisfying the physical constraints. The external force driving the system stays smooth throughout. Mathematically, the acceleration, pressure, momentum, and viscosity terms in the equations each grow large but cancel in a precise way, allowing the singularity to form without an infinite external input.

The result resolves statements C and D in the official Millennium Prize formulation.

The System

The model used was an internal OpenAI system described as significantly more capable than GPT-6 Astra. Training began August 28 and is ongoing. OpenAI did not name the model or announce a release date.

The solving effort launched September 1, prompted by rumors that two Millennium Prize Problems had been resolved independently. OpenAI ran a multi-agent system of coordinating groups against all remaining open Millennium Problems simultaneously.

The Navier-Stokes group involved approximately 10,000 concurrent agents. Across all attempted problems, agents sent 4.9 million messages and used about 300 billion output tokens. The agents reached their resolution on September 5 — 88 hours after launch.

Before focusing on Navier-Stokes, the same system also resolved the Euler regularity problem: the unforced version of the blow-up question for Navier-Stokes with the viscosity term removed. About 100 agents working 50 hours produced that result. The Euler solution then informed the Navier-Stokes effort, with OpenAI using Codex to consolidate insights across agent groups.

Context

Only one Millennium Prize Problem had previously been solved: the Poincaré conjecture, proved by Grigori Perelman in 2003 and confirmed in 2006. He declined the $1 million prize.

The practical consequence of the Navier-Stokes result is conceptual rather than engineering-critical. Real fluids are made of molecules and do not actually move infinitely fast — the singularity marks a breakdown in the continuum model, not a violation of physics. Computational fluid dynamics for aircraft design and weather forecasting uses the equations on discrete grids where the issue does not arise. The proof resolves whether the mathematical idealisation of a fluid has an inherent limit.

OpenAI stated that mathematical disputes remain ongoing. Coverage from wccftech noted that two mathematicians working on the problem are contesting the result. The Lean formalization is available alongside the proof writeup.

Key Numbers

• Agents on Navier-Stokes group: ~10,000 concurrent
• Time to resolution: 88 hours (Sept 1 → Sept 5)
• Messages sent across all problems: 4.9 million
• Output tokens used: ~300 billion
• Lean verification time: 17 hours via GPT-6 Astra
• Euler regularity: ~100 agents, ~50 hours (solved en route)
• Model: Internal, unnamed, more capable than GPT-6 Astra, training ongoing since Aug 28