OpenAI's 10,000-agent proof is the headline. The harder story is what's underneath it.
On September 8, OpenAI said an unreleased internal model, running roughly 10,000 coordinating agents over 88 hours, produced a proof that forced 3D Navier-Stokes equations can develop a finite-time singularity. GPT-6 Astra then spent 17 hours formalizing the result in Lean.
Here's the part most coverage skipped. Clay's Millennium Prize accepts two kinds of answers: prove smooth solutions always exist, or construct forcing that breaks that smoothness in finite time. OpenAI solved the second version and said plainly it isn't claiming the $1M prize. Clay still listed the problem unsolved the next day.
Then a credit dispute surfaced. NYU mathematician Tristan Buckmaster says he and Anthropic researcher Levent Alpöge reached a related Euler result on August 15 using Claude, Codex, and GPT-5.6 Sol. He grew suspicious once he learned OpenAI used the same uncommon “smooth forcing” approach, and asked whether OpenAI's model had touched their private Codex sessions. OpenAI said no user data was queried; Buckmaster says he never got clarity on training data. He also says OpenAI proposed he publish alone, without Alpöge as coauthor, which OpenAI researcher Sébastien Bubeck disputes.
OpenAI's own statement: no one saw the pair's results before they went public, but it can't fully rule out that de-identified usage data helped train the model somehow.
The more interesting story might not be the proof itself. It's that research at this scale now requires what production inference requires: decomposing tasks across thousands of agents, consolidating results, and knowing exactly what your model saw.