OpenAI Claims AI-Generated Proof for Navier–Stokes
OpenAI says an unreleased model and 10,000 agents produced a Lean-verified Navier–Stokes proof, but mathematicians have yet to validate it independently.
OpenAI says an internal system has produced a proposed solution to the Navier–Stokes existence and smoothness problem, one of the seven Clay Millennium Prize problems. The company says the result establishes that a smooth three-dimensional fluid can develop a finite-time singularity while retaining finite energy.
A large coordinated search
The proof was generated by a model OpenAI describes as significantly more capable than GPT-6 Astra. According to the company’s account, roughly 10,000 software agents worked concurrently for about 88 hours, using cached web material, code execution and formalisation tools. OpenAI also released a Lean version intended to make the argument machine-checkable.
The reported method focuses on a vortex that spirals inward while becoming increasingly elongated. The system first found a related blow-up construction for the Euler equations before extending the approach to the Navier–Stokes setting. That distinction matters because the Clay problem concerns the equations with viscosity and requires a globally valid mathematical argument, not merely a convincing numerical experiment.
The proof is not yet a prize
OpenAI has called the result a solution, but the mathematical community has not independently certified it. External scrutiny is especially important because the announcement followed reports that NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had made closely related progress. OpenAI denies using their private work, while acknowledging that it cannot completely rule out indirect influence from public or model-accessible material.
The significance is therefore twofold. If the proof survives review, it would be the first major Millennium problem resolved with decisive assistance from an AI system and a striking demonstration of long-horizon agent coordination. If it fails, the episode will still expose how difficult it is to assess frontier-model claims when the producing lab controls the model, compute trail and initial presentation.
Uncle Cat take
The 10,000-agent figure is less important than the missing independent referee: until mathematicians verify the Lean proof, this is an extraordinary claim, not a solved prize problem.