The humans chose the problem, the agents finished the task

On 1 September OpenAI set its new internal AI model on the six remaining Millennium Prize Problems, some of the most famous open questions in mathematics. About 88 hours later, according to the company, around 10,000 AI agents running at the same time had arrived at a solution to the Navier-Stokes problem, which had been open for close to 90 years. The result is not yet established as an accepted solution. But the way it was produced is already interesting.

It was the AI agents that did the actual problem solving. During the run they sent 2.7 million messages and produced around 130 billion tokens. When the work was done, OpenAI used GPT-6 Astra for a further 17 hours to rewrite the result in Lean, a system that can check that the steps in a mathematical argument follow logically from one another. The model that did the main work is not available to the public. OpenAI describes it as considerably stronger than Astra and says it is still being trained.

The humans' role lay mainly in deciding what the systems should work on. OpenAI's researchers formulated different variants of the problem and handed them out to groups of agents. They also let a smaller number of agents work on a simpler, related problem. When just under a hundred agents managed to solve it after about 50 hours, the researchers moved more resources to Navier-Stokes and let the other agents build on the result.

So the humans chose the direction and allocated the resources. The AI systems could do a large part of the search for the solution.

"Those who do not have access to the tools will fall behind in research," Svante Linusson, professor of mathematics at KTH Royal Institute of Technology, tells Dagens Nyheter.

Tristan Buckmaster at New York University and Levent Alpöge at Anthropic had been working on a related problem for much of the past year. During August they used Claude and Codex, among other tools, and by 22 August they had reached a result that had also been checked in Lean.

A researcher can have Claude or Codex as support in their work. OpenAI can instead put a stronger internal model and thousands of agents on the same problem at once. The difference is therefore not only about who has access to AI, but about what kind of research the tools make possible.

Whether OpenAI's solution holds remains to be seen. The proof and the Lean files are public, but the result has not yet had the broad independent review it takes to count as established. That does not change what the attempt already shows: the line for what can be handed over to AI in research has moved.