OpenAI turns Navier-Stokes into a stress test for agent-driven science

The question has resisted mathematicians for nearly ninety years. The equations describe everything from the flow around a wing to blood circulation and weather forecasting. The challenge was not to calculate a specific case, but to prove whether a smooth solution in three dimensions always remains smooth or can break down. OpenAI claims to have constructed a forced vortex that elongates and concentrates until its velocity grows without bound, while the external force remains smooth and the total energy remains finite. This would satisfy formulations C and D of the official problem statement: a refutation of universal smoothness, not an incremental improvement of a numerical method.

The process matters as much as the theorem. The company says the decisive group assembled on the order of 10,000 concurrent agents. It reached the result after about 88 hours and used another 17 to formalise and verify it with GPT-6 Astra. For Navier-Stokes, 2.7 million messages were exchanged and approximately 130 billion output tokens were generated. Codex consolidated intermediate ideas between groups exploring different approaches. This architecture resembles less an augmented individual researcher and more a temporary organisation: parallel search, specialisation, shared memory, cross-review, and a final verifiable artefact.

Lean raises the bar, but it does not eliminate mathematical judgement. A formal checker can verify that each step follows from declared rules; it does not in itself guarantee that the formal statement accurately represents the problem, that the physical assumptions are the required ones, or that overly strong axioms have not been introduced. The review of the analytical argument, its readability, and independent reproduction also remain. OpenAI is not claiming the one-million-dollar prize, and recognition does not depend on a corporate memo. The community will have to study the proof, detect potential modelling gaps, and decide if the result unambiguously fits within the Clay framework.

The scale also opens up an economic discussion. One hundred and thirty billion tokens, thousands of simultaneous processes, and an internal model more capable than the public product do not yet constitute a recipe available to a university or a company. OpenAI does not publish the computational cost, energy consumption, rate of failed attempts, or sufficient detail to repeat the orchestration. The relevant data for a lab is not that many agents can talk to each other, but how much it costs to obtain a valid claim, what part can be audited, and how many negative results or abandoned paths were left out of the report.

There is a second layer of governance. The project began after rumours of concurrent progress by Levent Alpöge and Tristan Buckmaster. On 10 September, OpenAI added an update to state that prompts Buckmaster had submitted to Codex in previous months could not have influenced the system, not even through training, and that the proofs differ. This clarification points to a problem that will recur: when scientific assistants receive unpublished work, a lab must demonstrate data separation, training traceability, intellectual priority, and the absence of leaks, not just claim that the model had an original idea.

For R&D organisations, the practical lesson is to start with work that has explicit verification: formal proofs, counter-example searches, unit reviews, hypothesis consistency, or calculation reproduction. Each agent needs a budget, a permitted source, a stopping condition, and a dependency log. The result must end in an artefact that a person or an independent tool can review. It is now worth monitoring the reaction from specialists, formal acceptance, the model's publication, and the real possibility of reproducing the coordination. If the proof holds up, the milestone will not be that an AI answered a maths problem; it will be that an agent infrastructure produced science that can be interrogated step by step.