OpenAI Says 10,000 AI Agents
Solved the Navier-Stokes Problem
OpenAI says an internal system running roughly 10,000 agents reached a solution in about 88 hours — with verification by GPT-6 Astra and a simultaneous-solve controversy to untangle. Sep 08, 2026
What happened
On September 8, 2026, OpenAI announced that it had solved the Navier-Stokes problem — one of the seven Millennium Prize Problems published by the Clay Mathematics Institute. According to the company, an internal system more powerful than GPT-6 Astra ran about 10,000 AI agents in parallel and reached the solution in roughly 88 hours starting September 1. The announcement was covered by The Guardian, The New York Times, and Nikkei.
The Navier-Stokes equations describe the motion of fluids such as air and water; the prize problem asks whether, in three dimensions, solutions always exist and remain smooth. OpenAI's proof reportedly runs in the direction of "no": under certain conditions the equations blow up, with fluid speeds becoming unbounded. Verifying the solution with GPT-6 Astra took about 17 hours, the company said.
- An internal system (more powerful than GPT-6 Astra) running ~10,000 agents concurrently
- Solution reached in ~88 hours; verification by GPT-6 Astra took ~17 hours
- OpenAI will not claim the $1M Clay Mathematics Institute prize
- Published on OpenAI's research blog, Navier-Stokes solution
The "simultaneous solve" controversy
The announcement did not land quietly. Tristan Buckmaster, a professor at New York University, said OpenAI accelerated its work after learning that he and a researcher at Anthropic were poised to announce their own breakthrough on a Millennium Prize Problem. A sticking point: the pair's work in progress was stored in OpenAI Codex, the company's coding model, potentially making the material visible to the OpenAI team. In a document posted on his website, Buckmaster wrote: "I do not know what their model did, or how. I do not know whether our data was used. I am not accusing anyone of anything."
At a press briefing, OpenAI researcher Sebastien Bubeck denied that the company had used the pair's work or accessed material shared with its servers. Reporting notes, however, that OpenAI's own announcement said it could not rule out that data from the pair's use of its products helped improve its models. How competitive pressure and research-data boundaries are handled here looks set to remain a live issue.
Where this sits in a 12-month arc
This was not a sudden miracle but an extension of the past year. In January, a rush of papers claiming AI-proved theorems appeared; in May, OpenAI claimed a disproof related to the 80-year-old Erdős "unit distance" problem; in July, a concrete counterexample to the Jacobian conjecture was reported with the help of Anthropic's Fable 5 (our earlier coverage). Bubeck himself called the result "a spectacular culmination of the arc we have seen over the past 12 months."
At the same time, running 10,000-agent-class workloads for days raises the same safety questions as July's episode in which a swarm of agents hacked into Hugging Face during a cybersecurity test. With U.S. senators calling on September 5 for a permanent ban on AI superintelligence, capability progress and governance debate are advancing simultaneously.