Fluid dynamics has had an open wound for roughly 90 years. The question of whether smooth three-dimensional fluid motion can mathematically break down, producing what’s called a singularity, has resisted every attempt at resolution since Jean Leray first prodded at it in 1934. OpenAI just closed it. Not with a team of mathematicians, but with an internal AI model that its own researchers describe as significantly more capable than GPT-6 Astra.
According to OpenAI, the system produced both a full analytical proof and a Lean formalization showing that a smooth fluid initially at rest can develop a singularity in finite time. That’s the kind of result mathematicians call a blowup: fluid velocity growing without bound within a finite window, even as energy stays finite throughout. The Clay Mathematics Institute had listed this as one of seven Millennium Prize Problems back in 2000, each carrying a $1 million prize. This is the first confirmed resolution of any of them by an AI system.
What the proof actually shows
The Navier-Stokes equations describe how fluids move. They’re used in aircraft design, weather forecasting, and modeling blood flow. The core question was whether the continuum approximation these equations rely on can break down entirely, even when motion starts smoothly and viscosity, which tends to dampen fluid motion, is present.
The solution takes the form of a vortex that spirals inward while stretching axially, a spaghettification of sorts. The central region shrinks while accelerating in a way that keeps total energy finite. What makes this mathematically hard is that the breakdown has to emerge from the fluid’s own motion, not from an externally applied infinite force. The terms describing acceleration, pressure gradients, momentum transfer, and viscosity must simultaneously grow large while canceling each other with precision, leaving only a smooth external force even as velocity diverges.
How the system was built and run
OpenAI started training the internal model on August 28. By September 1, after hearing rumors that two Millennium Prize problems may have been independently resolved, they launched a multiagent effort across all open Prize problems. The system coordinating the work had access to a cached version of the internet and code execution tools.
The numbers involved are significant:
- Around 10,000 concurrent agents worked on the Navier-Stokes problem specifically
- The agents exchanged 2.7 million messages and used approximately 130 billion output tokens for this problem alone
- The full resolution took about 88 hours from the first agent launch on September 1 to the answer on September 5
- Lean formalization and verification took an additional 17 hours, handled by GPT-6 Astra
Before committing resources to Navier-Stokes, the system also resolved a related problem: the regularity question for the Euler equations, which are essentially Navier-Stokes without the viscosity term. That result came from roughly 100 agents working together for about 50 hours. Seeing that success is what led OpenAI to redirect its agents toward the harder Navier-Stokes problem.
Why this matters beyond the math prize
OpenAI is making a specific point here about AI capability trajectory. The paper states directly that the internal model used is significantly more capable than GPT-6 Astra, and the company says it’s sharing this to inform the public about the pace of progress. That’s a deliberate signal, not just a research publication.
It’s also worth noting that concurrent independent work appears to have been underway. OpenAI mentions Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a math professor at NYU, as researchers whose work may have been the source of the September 1 rumors. So this is not an isolated AI result emerging from a vacuum. But an AI system resolving a 90-year-old Millennium Prize problem in under four days of compute time is a meaningful data point about where frontier models are heading, and how fast.
