Navier-Stokes proof: OpenAI’s 10,000-agent sprint claims 90-year math problem solved in Lean
Navier-Stokes proof headline: OpenAI announced on September 9, 2026, that an internal coordination of roughly 10,000 concurrent AI agents produced a Navier-Stokes proof for the long-standing Clay Millennium Prize problem, publishing both the mathematical argument and a Lean formalization in tandem. The run consumed 88 hours, approximately 2.7 million agent messages and an estimated 130 billion output tokens, with compute costs measured in the millions of dollars, according to a Benzinga report.
The agent swarm and the unreleased model
The work was carried out on an unreleased internal model that the company described as significantly more capable than GPT-6 Astra, with agents collaborating rather than operating in isolation. OpenAI positioned the simultaneous publication of a human-readable mathematical argument and a Lean formalization as a deliberate attempt to give reviewers two independent ways to audit the same Navier-Stokes proof, rather than relying on either one alone.
Altman: “one of the most amazing moments”
Chief Executive Sam Altman said the announcement ranked among the most memorable moments in OpenAI’s history, language he repeated across public posts and internal remarks covered in the Benzinga report. Co-founder Greg Brockman, in related coverage cited by the outlet, welcomed readers to what he called the AGI era, framing the mathematics result as a watershed for the lab’s broader research direction.
A credit dispute hours before the announcement
About 12 hours before OpenAI went public with its result, NYU’s Tristan Buckmaster and Levent Alpoge of Anthropic posted their own solutions to closely related fluid equations, an announcement that briefly threatened to overshadow the lab’s work. OpenAI said it had first heard a rumor about the pair’s progress on September 1, completed its own work on September 6, and proposed a joint announcement once the parallel timing became clear, only to later learn that the Buckmaster-Alpoge result covered the forced Euler equations rather than the full Navier-Stokes problem the Clay Millennium Prize targets.
Lean formalization published alongside the mathematics
The publication set released a Lean formalization of the Navier-Stokes proof alongside the narrative mathematics, treating the formal proof as a primary deliverable rather than an appendix. OpenAI also said it offered to share both its proof and the prompts used by the agent swarm with the Buckmaster-Alpoge team, an unusual step in a millennium-priority race where credit attribution is typically contested.
The prize, the cost and unresolved questions
Smooth solutions to the three-dimensional Navier-Stokes equations are one of seven Clay Millennium Prize Problems, each carrying a $1 million award, a status that sets a high external bar for any claimant. OpenAI acknowledged that it could not completely rule out anonymized user data contributing to improvements in the model behind the run, a disclosure that is likely to draw scrutiny as the formal review proceeds. Even with the Lean artifacts in hand, the broader mathematics community will judge the Navier-Stokes proof on whether the argument stands independently of the agent system that produced it.
Why this Navier-Stokes proof attempt will be judged differently
Past claims on the problem have largely been individual efforts that ended in partial results or, in some cases, retracted manuscripts, which is why the scale of this run is drawing both praise and skepticism. With a Lean companion and a public prompt exchange on offer, the company is treating the Navier-Stokes proof as a transparent artifact rather than a sealed result, leaving the final verdict to outside reviewers, the Clay Institute and time.
The Navier-Stokes proof run also surfaced a fresh debate inside the mathematics community about how credit should be assigned when a discovery is co-produced by humans and machines. OpenAI has argued that its internal model produced a Lean-checkable argument from scratch, while Buckmaster and Alpoge have noted that nearly a year of working with public Claude and Codex releases laid conceptual groundwork that the 10,000-agent sprint was able to exploit. The Clay Mathematics Institute has not commented publicly on whether the OpenAI submission qualifies for the $1 million prize, and the institute’s traditional rules require human authorship. That uncertainty has not slowed attention to the proof itself, which mathematicians at several US universities say they intend to audit step by step over the coming weeks.
The architecture OpenAI used for the Navier-Stokes proof run is also notable for what it consumed. Thea Dreisinger, an applied mathematician at UC Berkeley who has worked on Lean formalizations, told the Century Report that the 130 billion token figure is consistent with what a coordinated multi-agent search would generate when verifying intermediate steps and rewriting failed attempts. Dreisinger said the model’s ability to spot its own dead ends and pivot was what distinguished the run from earlier attempts at automated theorem proving. If the proof holds up under independent review, it will set a precedent for the kind of mathematics that frontier models can tackle end-to-end without human hand-holding, and it will reshape how research universities allocate graduate-level compute between human theorists and agentic systems they supervise.