Analysis
OpenAI said September 8 that mathematicians at the company, directing 10,000 autonomous AI agents running on an internal, unreleased model, found a 'singularity' in the three-dimensional Navier-Stokes equations -- resolving one of the six Millennium Prize Problems the Clay Mathematics Institute posed in 2000, Quanta Magazine reported. The model took 88 hours to reach the result, CNN reported separately, and OpenAI said it does not intend to claim the $1 million Clay Institute prize for the result.
How the result was verified
The result has been formally checked in Lean, a machine-verification proof assistant that lets other mathematicians independently confirm each logical step without having to trust OpenAI's own account of the proof -- a materially stronger form of verification than a natural-language proof reviewed by a handful of referees, Nature reported, and the same tool OpenAI has increasingly relied on to defend its recent string of mathematical claims against skepticism.
That skepticism has been substantial: OpenAI's own work on this problem began September 1 after the company heard what it described as a rumor that another Millennium Prize problem had been separately resolved -- work that turned out to reference mathematicians Levent Alpoge and Tristan Buckmaster, who published their own Navier-Stokes-adjacent result around the same time through conventional peer-reviewed channels rather than a 10,000-agent compute run. Pulse has separately covered a widening dispute over whether OpenAI's Astra model's other claimed math breakthroughs properly credited researchers whose prior conversations with ChatGPT may have fed into results the company later presented as autonomous.
What Lean verification does and doesn't settle
A Lean-checked proof is a genuinely different standard of confidence than OpenAI's now-disputed Astra manuscript, which claimed ten solved open problems without full formal verification and has drawn credit disputes from at least two named mathematicians. But formal verification confirms the logic of a proof is internally consistent -- it doesn't by itself confirm the proof represents genuinely novel mathematical insight versus, in principle, a restatement or minor extension of ideas already present in training data the model was never fully audited against. Traditional peer review, still ongoing for this result as of publication, is the process that would settle that harder question -- and it moves on a timescale of months, not the 88 hours OpenAI's compute run took to generate the claimed proof itself.
The result nonetheless lands as one of the more credible AI mathematics claims of 2026, precisely because Lean verification removes the single biggest objection critics raised against the company's less rigorously checked Astra manuscript. Whether 10,000-agent, formally-verified proof generation becomes a repeatable process for OpenAI -- rather than a one-off applied to a famous, well-defined problem with an unusually clean formalization -- is the open question the next several months of peer review, and any follow-up results, will need to answer.
