Build1 publisher3 min readPublished
OpenAI gated an 88-hour, 10,000-agent proof search on a 17-hour Lean check
OpenAI's write-up gives the token counts, the agent count and the verification time for its Navier-Stokes result. The verification time is the number that decides whether the method transfers to anyone else's workload.
The Engineer · Build desk
What happened
- OpenAI's technical write-up says the Navier-Stokes run used roughly 10,000 concurrent agents, 2.7 million messages and 130 billion output tokens across 88 hours, plus about 17 hours of checking in Lean.
- OpenAI says the result establishes finite-time blowup for 3D incompressible Navier-Stokes under a smooth force from smooth initial conditions, and it does not itself claim the Clay prize.
- Andrew Curran says OpenAI told The New York Times it had made "substantial progress" on another Millennium Prize problem within five days of the Navier-Stokes push and was preparing to announce more.
- The New York Times reported that NYU mathematician Tristan Buckmaster had been pursuing the Navier-Stokes line before OpenAI published first, turning the result into a fight about priority.
- Anthropic published a September threat-intelligence report covering misuse it disrupted from December 2025 through August 2026, across cyber operations, surveillance, influence, weapons, scams and distillation.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- capability A cheap and total checker is what makes it rational to run 10,000 agents unattended, because the acceptance decision costs about a fifth of the search and needs nobody to read the transcripts.
- cost Anyone copying the method is buying compute they have to be willing to write off, since Armstrong's account puts the bill in the millions and has humans doing little except moving capacity between paths.
- constraint Work whose acceptance test is a human reading the output gets nothing from this fan-out, because the parallelism multiplies the review queue instead of shortening it.
- exposure Mathematicians who put unpublished lines of attack into a lab's products now have to litigate whether their prompts reached training, with OpenAI saying Buckmaster's two months of Codex prompts could not have.
The 17 hours in Lean is the part of that run worth copying. Formal checking cost about a fifth of the search time [1], and the verdict is binary, since the accompanying repository holds a machine-checkable proof that either compiles or does not [2]. Nobody has to read 2.7 million messages to find out whether the search worked.
Divide the tokens by the wall clock and the search itself looks ordinary. 130 billion output tokens over 88 hours is about 410,000 tokens per second in aggregate, which across 10,000 agents is roughly 41 tokens per second each [2]. That is one unremarkable stream per agent, held for the full 88 hours. The scarce inputs were parallel capacity and a checker that could clear the backlog.
One figure needs care. 130 billion output tokens across 2.7 million messages averages about 48,000 output tokens per message [4], which is not a chat turn. Either the messages carry large artifacts, or the two counters are counting different things.
Against the earlier Euler work, the change is fan-out. Roughly 100 agents for about 50 hours is 5,000 agent-hours; 10,000 agents for 88 hours is 880,000, about 176 times as much [3]. Evan Armstrong's account says humans mainly reallocated compute across promising paths, and that Noam Brown put the bill in the millions [7].
Armstrong frames the mechanism as "whatever compute can check, the labs can conquer" [5]. The checker is the condition. Three properties have to hold before fan-out pays: the check is cheap relative to the search, it is total, and its verdict is trusted without anyone reading the work. Lean has all three. A test suite has the first two, and the third only as far as its coverage goes. A human reviewer has none of them, and 10,000 agents behind a human acceptance gate produce 10,000 things to read.
Curran's follow-on claim rests on less. He says the leading rumor is the Hodge Conjecture and speculates that the unreleased post-Astra model is called "Aeon," and OpenAI has confirmed neither detail [9].
Andreas Thom argues that researchers cannot trust OpenAI with unpublished mathematics, after his own ChatGPT conversations around an expander-matching problem [11]. Valerio Capraro amplified the allegation, arguing that if the accounts involving Levent Alpöge, Tristan Buckmaster and Thom are substantiated, unpublished human work may have been absorbed into a model and presented as an AI breakthrough [17]. OpenAI's roon, who says he was on the first announcement call, said the team moved from lawyerly hedging about opt-in Codex data to saying the chance those prompts entered training was zero [13].
Alongside the mathematics, the digest lists a California audit regime, Anthropic's widest misuse report, Codex turned into an agent API and a Wall Street Astra workspace as one theme [16]. For three of those it supplies the name. The Anthropic report carries detail: the New York Times focused on five biological cases that could support weapons development, and Anthropic stressed it was not asserting the scientists intended harm [15]. This material does not say what California audits, or what the Codex agent API permits by default, so the proposition that deployers are absorbing the new risk cannot be tested against it.
What to watch
- Whether the next Millennium result ships with a machine-checkable repository or only a claim of substantial progress.
- Whether OpenAI names the internal model above GPT-6 Astra, which would test Curran's "Aeon" speculation.
- Whether Buckmaster or Alpöge publish their own timeline of the priority dispute.