On the Navier–Stokes Millennium Prize Problem

OpenAI’s technical post asserts a solution to the Navier–Stokes existence and smoothness problem: an initially smooth fluid at rest, under a smooth force with finite energy throughout, can develop a singularity in finite time. The company is publishing a write-up plus a Lean formalization, and it frames the work as evidence of pace rather than a prize grab.

Methodologically this was an industrial agent campaign. Training of a new internal math-strong model began August 28. After September 1 rumors that Millennium problems might be falling, OpenAI launched coordinating agents with tools (cached web, code execution), group chat, and strict evaluation safeguards. The Navier–Stokes group ran on the order of 10,000 concurrent agents; resolution came September 5 after about 88 hours, with 17 more hours of Lean formalization via GPT-6 Astra.

Token burn was enormous—about 130 billion output tokens and 2.7 million messages on Navier–Stokes alone (4.9 million messages / ~300 billion output tokens across attempted problems).

Signal Desk’s read: treat this as a systems result first—swarm search plus formal verification—and a mathematics result second until the community stress-tests the Lean artifact. OpenAI explicitly routed separate agent groups onto statements A/B (proof-shaped) and C/D (disproof-shaped). Agents first resolved an “easier” unforced Euler regularity blowup (~100 agents, ~50 hours), then management concentrated resources on Navier–Stokes and cross-pollinated insights with Codex.

The post’s careful diplomacy toward Levent Alpöge and Tristan Buckmaster is load-bearing. OpenAI says it reached out after Lean verification (September 6) believing they might also have Navier–Stokes, then learned they had forced Euler; it recognizes their priority there, notes the Euler variants differ (forced vs unforced), and states researchers and agents did not see the pair’s work before public release—while still unable to rule out de-identified product-derived training signal. That caveat is the honesty tax of running customer agents and research agents inside one corporate boundary.

What is being over-claimed in the broader discourse is “AI solved a Millennium Problem, prize pending.” OpenAI itself says it does not intend to claim the prize and casts the release as a progress snapshot from a model still training. The under-claimed piece is operational: prize-level math is now something a lab can attempt with a weekend of agent orchestration if it is willing to spend at this token scale.

The likelier lasting effect is a new verification culture—Lean or bust—for any AI-announced frontier proof, plus sharper norms about rumor-triggered compute races.

The Clay Mathematics Institute listed Navier–Stokes among seven Millennium Prize Problems in 2000. Jean Leray’s 1934 work established generalized solutions; smoothness versus singularity for 3D incompressible flow remained open. OpenAI describes the singular solution as an inward-spiraling, elongating vortex that keeps energy finite while velocity unbounded.