On the Navier–Stokes Millennium Prize Problem

TL;DR

OpenAI announces that a coordinated swarm of ~10,000 AI agents, driven by an internal model described as “significantly more capable than GPT-6 Astra,” produced a proposed resolution to the Navier–Stokes existence and smoothness Millennium Prize Problem: a finite-time blowup construction for the 3D incompressible Navier–Stokes equations under smooth forcing, with a self-similar vortex that tightens indefinitely while keeping total energy bounded. The release ships a 165-page manuscript plus a Lean formalization that mechanically verifies the argument end-to-end. OpenAI says the winning agent group reached its result 88 hours after launch (2.7M messages, ~130B output tokens for this problem; 4.9M messages / ~300B tokens across all attempted problems), with an additional 17 hours of Lean formalization aided by GPT-6 Astra. OpenAI says it will not claim the $1M Clay Prize. The release lands amid a public priority dispute with Levent Alpöge (Anthropic) and Tristan Buckmaster (NYU); see Statement on OpenAI Navier–Stokes priority dispute (Buckmaster, 2026-09-08).

Key claims

The system’s proof construction is finite-time blowup: an initially smooth fluid at rest, subjected to a smooth external force, develops a singularity in finite time while total energy remains bounded — the resolution corresponds to options (c) and (d) in Fefferman’s official Clay problem statement (smooth forcing on ℝ³ / 𝕋³) [§Result].
Agent swarm reached its resolution ~88 hours after launch; ~10,000 coordinating agents were deployed with an internal model “significantly more capable than GPT-6 Astra” driving them [§Method].
Cost: 2.7 million agent-to-agent messages and ~130 billion output tokens for the Navier–Stokes problem specifically; 4.9 million messages and ~300 billion output tokens across all problems the swarm attempted [§Compute].
Lean formalization and end-to-end verification took an additional 17 hours, performed with GPT-6 Astra assistance — trusting the result therefore reduces to trusting the Lean kernel and its axioms, not the model’s natural-language reasoning [§Formalization].
An unforced-Euler intermediate result (finite-time blowup for the inviscid limit) was produced first by ~100 agents in ~50 hours; this success redirected compute toward the full Navier–Stokes problem, with the Euler result fed back into remaining agent groups [§Method].
Priority acknowledgement: OpenAI says its effort began on September 1 after hearing a rumor which it later realized traced to Alpöge (Anthropic) and Buckmaster (NYU); upon completing its own proof and Lean verification on September 6, it reached out to offer a concurrent release, at which point it learned their work addressed the related-but-distinct forced Euler problem rather than Navier–Stokes [§Timeline].
OpenAI states it “cannot rule out that de-identified data from their use of OpenAI products was used” in training, but says agents did not access the researchers’ work through any means before public release [§Data].
OpenAI will not claim the $1M Clay Prize [§Prize].

Method

The release describes a group of agents working in parallel, coordinated through an internal OpenAI system built on a next-generation model that is not the publicly-released Astra. The workflow is described as: (1) launch multiple agent groups in parallel on the problem, (2) accept intermediate progress (unforced Euler blowup) and redirect remaining compute toward the harder target, (3) use Codex to consolidate insights across agent groups, (4) generate a natural-language manuscript, (5) translate to Lean and iterate against the compiler until every step type-checks. The construction geometry — a self-similar inward-spiraling vortex with bounded energy but unbounded velocity at the singular time — sits in the line of work started by Diego Córdoba and Luis Martínez-Zoroa, whose smooth-force blowup program was the analytic route Buckmaster/Alpöge were pursuing in parallel (see Statement on OpenAI Navier–Stokes priority dispute (Buckmaster, 2026-09-08)).

Results

One Millennium Prize Problem addressed (Navier–Stokes existence and smoothness under smooth forcing).
165-page manuscript + Lean formalization released publicly [§Release].
Intermediate results: finite-time blowup for unforced 3D Euler was produced along the way by ~100 agents in ~50 hours [§Method].
Compute totals: 4.9M messages / ~300B output tokens across all attempted problems; 2.7M messages / ~130B tokens for Navier–Stokes alone; ~88 hours wall-clock for the winning group + 17 hours for Lean [§Compute].

Why it’s interesting

Two things sit on top of each other here. First, the technical claim: if the Lean formalization holds up under independent inspection, this is the first machine-verified resolution of one of the Millennium Problems, produced by a system that is a direct continuation of the AI-for-AI-research scaffold pattern the wiki has been tracking — the ~10,000-agent swarm coordinated by an internal frontier model is a step-function extension of the ASI-ARCH / FARS / Karpathy autoresearch pattern already indexed at AI-for-AI Research, operating at a granularity (open mathematics problem) that until now had a single filed instance (LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks for a Knuth subproblem). Second, the priority controversy: OpenAI’s own timeline — “our effort began on September 1 after hearing a rumor” — is a first-person confirmation of the trigger described in Statement on OpenAI Navier–Stokes priority dispute (Buckmaster, 2026-09-08), though the two accounts differ sharply on what happened next. The compute-cost anchor is worth noting: 300B output tokens across all attempted problems and ~ $millions in compute (per WSJ) sets a public price tag on frontier-lab autonomous-math effort that other filed AI-for-AI-research systems (Karpathy autoresearch at ~$ 300 GPU + $9 API; FARS at ~$ 1K/paper) do not approach.