OpenAI says an unreleased model coordinated about 10,000 AI agents for 88 hours to produce a proposed solution to the Navier-Stokes existence and smoothness problem, one of mathematics’ seven Millennium Prize Problems. The run used roughly 130 billion output tokens and generated 2.7 million messages, according to the company.
The paper argues that a smooth, three-dimensional fluid starting from rest can develop a singularity in finite time while its energy remains finite. In physical terms, the proposed construction is a vortex that spirals inward and accelerates without bound even though the external force remains smooth. Establishing that counterexample would resolve the problem in the negative.
OpenAI published an analytical paper and a formalization in Lean, software that checks whether every step follows from the encoded premises. The agents reached the result on September 5. GPT-6 Astra then spent another 17 hours formalizing and verifying it. A checked Lean file does not settle whether the formal statement perfectly matches the prize problem or establish acceptance among mathematicians.
The Clay Mathematics Institute will not consider a proposed solution until it appears in a qualifying outlet, at least two years have passed and it has gained general acceptance in the global mathematics community. OpenAI said it does not intend to claim the $1 million prize. Independent review of the newly published work is only beginning.
The experiment began September 1, four days after OpenAI started training the model that powered it. The company describes that system as significantly more capable than GPT-6 Astra and says its training is continuing. Agents could read a cached copy of the internet, run code and communicate within groups. After nearly 100 agents spent about 50 hours finding a related result for the unforced Euler equations, OpenAI shifted more agents to Navier-Stokes and used Codex to consolidate useful work across groups. A newer training checkpoint was introduced during the run.
The scale exceeds OpenAI’s earlier ten AI-generated mathematics results, which also came with Lean certificates. Across every problem attempted in the latest project, the agents produced 4.9 million messages and about 300 billion output tokens. OpenAI research chief Mark Chen told WIRED that the compute cost was in the millions of dollars. The company has separately projected roughly $750 billion in compute spending through 2030.
The announcement also carries a dispute over how the effort began. OpenAI says it acted after hearing rumors that NYU professor Tristan Buckmaster and Anthropic employee Levent Alpöge had made related progress. Buckmaster says OpenAI learned about their private work, accelerated its own effort and sought to influence how credit would be assigned. OpenAI says neither its researchers nor its agents saw their work before it was made public and disputes parts of Buckmaster’s account. It also says it cannot rule out the possibility that de-identified data from their use of its products had indirectly improved its models.
Sources: OpenAI, Clay Mathematics Institute, WIRED
–
By the Control Plane Editorial Team