OpenAI Claims 10,000 Agents Solved Navier-Stokes Millennium Problem in 88 Hours

On September 9, OpenAI announced that a large-scale agent swarm has produced a solution to the Navier–Stokes problem, one of the Millennium Prize Problems (whether three-dimensional smooth fluid motion can develop singularities/blow-up, open for roughly 90 years). The news was amplified by reposts from sama and others, drawing widespread attention.

Confirmed

• OpenAI claims it deployed roughly 10,000 concurrent agents running continuously for 88 hours, with the same safety safeguards (monitoring and isolation) as frontier evaluations applied throughout.
• The output was produced by a set of new internal models still in training; according to a technical note relayed by theo bearman, this model has been training since August 28 and is significantly stronger than GPT-6 Astra.
• Per run details compiled by danielmac8, the process output roughly 130 billion tokens; nrehiew's recap also mentions millions of messages and follow-up formalization work involving around 300 billion tokens.
• The timeline assembled by nrehiew shows: on August 15, Buckmaster and Alpoge obtained a solution for the forced Euler equations; work progressed around August 22; the agent swarm then closed in on the Navier–Stokes solution within 88 hours, with formalization completed using GPT-6 Astra.

Not Confirmed

• The above information comes from OpenAI's official claims and technical notes relayed by reposters; the correctness of the solution and whether it will be accepted by the mathematics community await independent verification.
• The specific name, architecture, and capability details of the new internal model have not been disclosed.

Why It Matters

• If the solution holds, it would be the first Millennium Prize Problem cracked by a large-scale agent swarm, marking substantive AI participation in frontier pure mathematics research.
• The scale—ten-thousand-agent collaboration, 88 hours, hundreds of billions of tokens—showcases the new research paradigm of "agent swarm + strong models"; SwarmOS developer bingxu also called it "a use case for the era of great discoveries."

2026-09-09 ~ 2026-09-09 · 14 related posts

• [source] OpenAI's NS formalization: 10k agents, 88 hours, ~130B output tokens — nrehiew_ · 2026-09-09
• [source] Agent swarm burns 300B tokens in 88 hours to crack Navier-Stokes, dev claims — bingxu_ · 2026-09-09
• [source] Navier-Stokes run details: 10k agents, 88 hours, 270M messages, 130B output tokens — daniel_mac8 · 2026-09-09
• OpenAI says its model solved Navier-Stokes in 88 hours using ~10,000 coordinating AI agents — HankYeomans · 2026-09-09
• OpenAI claims agent swarm proved 90-year-old Navier-Stokes Millennium Prize Problem — sama · 2026-09-09
• OpenAI Claims Navier-Stokes Millennium Problem Solved by 10,000-Agent Swarm — connoraxiotes · 2026-09-09
• OpenAI Claims Its Agents Produced a Proof of the Navier-Stokes Millennium Prize Problem — jonstephens85 · 2026-09-09
• OpenAI's unreleased post-GPT-6 model solves 90-year-old math problem in 88 hours — mrtac96 · 2026-09-09
• OpenAI says its Navier-Stokes project ran ~10,000 concurrent agents under strict safeguards — LatentSpaceLeaper · 2026-09-09
• 88 hours to crack a 90-year-old fluid problem: OpenAI agents tackle Navier-Stokes — mdancho84 · 2026-09-09
• OpenAI claims agent swarm solved the Navier-Stokes Millennium Prize Problem — austinc3301 · 2026-09-09

3 near-duplicate retellings: soumitrashukla9 · austinc3301 · mobav0