OpenAI on September 8 published a Lean-formalized proof of finite-time singularity in the three-dimensional Navier–Stokes equations, one of the Clay Mathematics Institute's $1 million Millennium Prize problems, and disclosed that the work was produced by an internal system it describes as "significantly more capable than GPT-6 Astra." The model, according to OpenAI's writeup, began training on August 28 and was still training at time of publication.

The mechanics are the story. Roughly 10,000 concurrent agents worked the Navier–Stokes group. They arrived at a resolution in 88 hours, with another 17 hours of Lean formalization run on GPT-6 Astra itself. The Navier–Stokes effort alone consumed about 130 billion output tokens and 2.7 million inter-agent messages; the full Millennium sweep across problems ran to 4.9 million messages and roughly 300 billion tokens. TechCrunch, citing OpenAI, pegs the week's compute at about $22.5 million at current Astra rates. The Washington Post characterizes it as "just days of effort."

The priority claim is already contested. Tristan Buckmaster of NYU and Levent Alpöge, affiliated with Anthropic, announced closely related proofs, including a Lean-verified Euler result, roughly 12 hours before OpenAI's post, per Quanta Magazine. Both efforts built on analytic "infinite cascade" techniques pioneered by Diego Córdoba and Luis Martínez-Zoroa. Buckmaster alleges to TechCrunch that OpenAI accelerated its own run after learning of his group's progress. OpenAI's post says no specific user data was accessed while acknowledging it "cannot rule out that de-identified data" from Codex usage improved its models.

Read that sequence carefully. It rhymes with the AlphaGo-versus-DeepMind-vs-Fan-Hui moment of 2016, where the institutional player timed disclosure to swallow the human milestone. What's new is the deployment shape.

OpenAI isn't selling a bigger single model. It's selling swarms. Fortune reports the company's internal research organization is already running 3.1 agent-workdays per human workday as of mid-August, with 90th-percentile researchers burning upwards of $7,000 a day in agent compute. GPT-6 Astra, covered in our review of its computer-use and agentic professional work, and Anthropic's parallel push on business-automation benchmarks with Fable 5.1, are now the floor, not the frontier.

For anyone whose customer-acquisition workflow depends on research, competitive mapping, prospect discovery, positioning, content strategy, the implication is legible. The work that a human analyst used to do in a week is now a $22.5 million question at Millennium-Prize depth and a rounding-error question at small-business depth. The gap between operators who have already wired agents into that workflow and operators who haven't is compounding with every release cycle. Astra was a preview. This is the shape of what compounds on top of it.

Sources

https://openai.com/index/navier-stokes-solution/
https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/
https://techcrunch.com/2026/09/08/openai-fought-dirty-on-career-making-math-problem-says-nyu-mathematician/
https://www.washingtonpost.com/technology/2026/09/09/openai-claims-it-solved-elusive-math-problem-with-1-million-prize/
https://fortune.com/2026/09/08/openai-rsi-progress-jakub-pachoki-warns-dangers-slowdown-safety-rules/