OpenAI has announced that an internal AI system has produced a solution to the Navier–Stokes existence and smoothness problem — one of the seven Millennium Prize Problems — marking what may be the most significant scientific result ever produced by an AI system.
The Problem
The Navier–Stokes equations describe fluid motion using Newton’s second law, treating fluids as continuous media. They govern aircraft design, weather forecasting, and blood flow studies. The open question: can smooth three-dimensional incompressible fluid motion develop a singularity — speeds growing without bound in finite time — despite viscosity’s smoothing effect?
The question dates to the 19th-century work of Navier and Stokes. Jean Leray proved generalized solutions exist in 1934, but smoothness remained unresolved. In 2000, the Clay Mathematics Institute named it one of seven Millennium Prize Problems, each worth $1 million.
The Result
The proof shows that an initially smooth fluid at rest, with a smooth applied force and finite energy, can develop a singularity in finite time — resolving statements “C” and “D” of the official problem formulation. The mechanism: a vortex that spirals inward and elongates “like spaghetti,” speeding up while its energy remains physically finite. The technical heart is that the equation’s terms — acceleration, pressure gradients, momentum transfer, viscosity — grow large yet cancel precisely, keeping the external force smooth even as velocity becomes unbounded.
Both a written proof and a Lean formalization are published.
How It Happened
Since August 28, OpenAI has been training a new internal model with “unprecedented performance” in mathematics. On September 1, after hearing rumors of Millennium Prize resolutions, the team launched a coordinated evaluation across all open Millennium Prize problems.
- Agent swarm architecture — groups of communicating agents with internet cache and code execution access; the winning Navier–Stokes group ran ~10,000 concurrent agents
- Variant coverage — separate groups explored problem formulations A, B, C, and D
- Cross-pollination — Codex consolidated insights across groups, feeding useful intermediate results back as new prompts
- The surprise — agents first resolved the unforced Euler regularity problem (~100 agents, 50 hours), which redirected the full effort to Navier–Stokes
- The cost — 88 hours to the proof; 2.7M messages and ~130B output tokens for Navier–Stokes alone (4.9M messages, 300B tokens across all problems); plus 17 hours of Lean verification via GPT-6 Astra
Concurrent Work
After completion, OpenAI learned the original rumor concerned Levent Alpöge (Anthropic) and Tristan Buckmaster (NYU), who had resolved the forced Euler problem. OpenAI credits their priority, notes the proofs differ significantly (their Euler result was forced; OpenAI’s was unforced), and states no specific user data was accessed — while acknowledging it “cannot rule out” that de-identified product usage data improved the models.
What It Means
OpenAI frames the result carefully: it will not claim the Millennium Prize, calling this “a snapshot in time of progress on AI development.” The announcement’s real payload is elsewhere — the existence of an internal model significantly beyond GPT-6 Astra, the effectiveness of massive agent coordination on open scientific problems, and what this implies about “the next period of AI progress.”
Coming one week after Astra’s launch — which itself produced new prime-gap proofs — the result suggests AI-assisted mathematics has crossed from “helping with calculations” to “producing frontier discoveries.” The Clay Institute’s prize committee now faces an unprecedented question: what happens to a $1 million prize when the solver is a machine that declines the award?