OpenAI researcher Noam Brown revealed that a system of 10,000 AI agents solved the Navier-Stokes existence and smoothness problem—one of the seven Millennium Prize Problems—using 130 billion tokens over 88 hours. Brown, a foundational contributor to OpenAI's o1 reasoning models and now working on multi-agent systems, detailed the achievement on a podcast published September 17, 2026.
The multi-agent approach parallelizes test-time compute rather than extending a single model's serial reasoning. Brown described it as less efficient than a single agent with full context but an effective way to scale inference when latency becomes a bottleneck. Published benchmarks show four agents completing tasks twice as fast at 2x cost, with 16 agents continuing the pattern at slightly lower efficiency. Math and web search tasks parallelize well; creative writing does not.
Confirmed
- 10,000 agents, 130 billion tokens, 88 hours to solve Navier-Stokes (a Millennium Prize Problem).
- Brown attributes the breakthrough primarily to a "very powerful" base model capable of long-horizon reasoning, not the multi-agent architecture itself—estimating multi-agent contributed less than 10% of the credit.
- Ultra Mode defaults to four agents; users can increase the count. Published scaling plots show sublinear speedup up to 16 agents.
- Training on verifiable synthetic problems generalized to the far more ambitious Millennium Prize Problem.
- Brown warns that as models improve, finding sufficiently hard training problems becomes a challenge—unlike systems with infinite self-play curricula.
- Dwarkesh Patel framed the 130-billion-token run as roughly 4,000 years of full-time human thinking compressed into 88 hours — a host conversion, not an OpenAI claim.
- Brown said his prior expectation put Millennium-Prize-level math closer to around 2028; he was surprised the Navier-Stokes run arrived sooner, and still credits the base model over multi-agent scaffolding (<10%).
Unknown
- How long a single agent would take to solve the same problem—no ablation experiment has been run at that scale.
- Whether the 10,000-agent configuration provided meaningful benefit over 1,000 or 100 agents. Brown calls the weekend run "one data point" and says thorough ablations are prohibitively expensive.
- Whether the result has been formally verified or submitted for the Clay Mathematics Institute's $1 million prize.
- Model version, hardware, or cost figures for the 130-billion-token run.
Our take
The Navier-Stokes result is a striking demonstration of inference-time scaling. But the real story is Brown's candor: OpenAI has one data point at 10,000 agents and no controlled comparisons at intermediate scales. Since Brown himself credits the base model with overwhelming weight, the 10,000-agent figure remains a capability ceiling until independent replication or systematic ablations exist. Related Brocker coverage: alignment thread from the same interview.
