OpenAI says 10,000 agents helped crack a 90-year-old math problem

01

WHAT ACTUALLY CHANGED

OpenAI released a proposed solution to the Navier–Stokes existence and smoothness problem on September 8. The problem asks whether the equations used to describe three-dimensional fluid motion must remain smooth forever when they start smoothly, or whether they can develop a singularity where velocity becomes unbounded in finite time.

The company's 166-page paper constructs the breakdown case. It describes a smooth fluid starting from rest with a smooth external force. A vortex contracts into an increasingly narrow region while its velocity rises without bound, even though the total kinetic energy remains finite. OpenAI says this establishes alternatives C and D in the official Millennium Prize formulation for ordinary space and a periodic three-dimensional torus.

The result came from a multi-agent system powered by an internal model that OpenAI says is significantly more capable than GPT-6 Astra. The successful effort used on the order of 10,000 concurrent agents. OpenAI reports that the agents reached the Navier–Stokes construction after about 88 hours, sent 2.7 million messages, and generated roughly 130 billion output tokens for that problem.

OpenAI also published a Lean 4 formalization. The public repository contains the Navier–Stokes and Euler certificates, build instructions, and a separate path for checking them with Comparator. OpenAI says GPT-6 Astra completed Lean formalization and verification during an additional 17 hours after the agents produced the mathematical result.

The recognition step has not happened. The Clay Mathematics Institute still lists Navier–Stokes among its unsolved problems. OpenAI says it does not intend to claim the $1 million prize. More importantly, a new proof this consequential needs broad review by mathematicians who can inspect both the ordinary argument and the exact statement encoded in the proof assistant.

02

WHY THIS MATTERS

If the proof survives review, it would resolve one of the best-known open questions in mathematics and show an AI system contributing to research far beyond checking a textbook exercise. Navier–Stokes equations appear in aircraft design, weather models, blood-flow studies, and many other places where fluids need to be described continuously.

The particular result does not mean airplanes suddenly stop obeying physics. The construction concerns whether the mathematical equations can produce a singularity under allowed conditions. A real fluid cannot achieve infinite speed, so the singularity would mark a boundary where this continuous mathematical description stops being sufficient.

The process may be as important as the theorem. OpenAI did not ask one chatbot to stare heroically at a blank page. Thousands of agents explored different formulations, shared intermediate insights, used tools, and were redirected toward a promising route. This looks less like one synthetic genius and more like an industrial research organization compressed into a few days.

Formalization supplies an unusually strong receipt. If Lean accepts the project, it has mechanically checked that the encoded conclusion follows from the encoded assumptions and definitions without missing algebraic steps. That is powerful, but it is not a magic authenticity stamp. People still need to verify that the formal statement really matches the Clay problem, that the definitions express the intended mathematics, and that no untrusted assumptions smuggled the answer into the system.

There is also a governance signal hiding inside the mathematics. OpenAI says the internal model is still training and already exceeds its newly released frontier model. A system capable of organizing 10,000-agent scientific campaigns creates obvious opportunities for discovery, along with difficult questions about compute access, priority, credit, disclosure, and how the rest of science gets enough time to check what arrives.

03

WHERE IT COULD HELP

• Use coordinated agents to explore many mathematical approaches in parallel
• Translate long analytical arguments into machine-checkable proofs
• Audit scientific claims by publishing source proofs and formal certificates
• Apply proof assistants to high-consequence engineering mathematics
• Study where large agent collectives accelerate research and where they create verification bottlenecks