Research index

As of 8 September 2026. OpenAI published a claimed solution. Clay has not awarded a prize. Independent verification has had hours, not years.

Executive abstract

OpenAI says an unnamed internal model, run as a swarm of about 10,000 coordinating agents, produced a 165-page proof and a Lean formalization that three-dimensional incompressible Navier–Stokes can develop unbounded velocity in finite time from rest, under a smooth external force, while kinetic energy stays bounded. The company maps this to Charles Fefferman’s official alternatives (C) and (D) — the breakdown doors that allow a smooth force — not to unforced global regularity (A) or (B). The constructed object is a contracting vortex that spirals inward and stretches like spaghetti; oscillatory corrections keep the leftover force smooth even as speed diverges.

That is important in mathematics because it would close one of Clay’s four written doors on a supercritical PDE that has been open since Leray (1934). It is not, on any honest one-year horizon, a new aircraft, a new climate model, or a turbulence theory. Engineers already compute these equations. The theorem is about whether the continuum model can fail. David Silvester’s “mathematical nicety” is the right practical counterweight.

What it unlocks, if it holds: a negative well-posedness theorem for the forced problem; a public Lean artifact; a demonstration that agent search plus a proof assistant can finish a Millennium-scale analysis construction; and a live test of Terence Tao’s warning that answers can arrive faster than understanding. What it does not unlock: unforced Navier–Stokes, a Clay cheque (OpenAI says it will not claim one), or a settled credit story with Tristan Buckmaster and Levent Alpöge, who the same day released forced blowup results for Euler, Boussinesq, and porous media.

Confirmed: the announcement, PDF, and GitHub repo exist; Fefferman’s (C)/(D) allow smooth force; OpenAI also claims unforced Euler blowup as a separate theorem. Unverified: that Lean matches Fefferman and that the analysis has no gap. Rumor/process: whether Codex sessions or rumor-steering chose the route.

Short answers

Question	Short answer
Did OpenAI solve “Navier–Stokes”?	They claimed Clay (C) and (D) (forced breakdown). Not (A)/(B) (unforced smoothness).
Is the million dollars paid?	No. OpenAI is not claiming it. Clay’s process has not started in any official sense.
Why does math care?	It is the flagship supercritical fluids PDE; a door closing is a foundational well-posedness fact.
Why does the rest of the world care?	The continuum model under CFD might fail; the method (agents + Lean) may travel. Wings do not change next week.
How?	Unreleased model > GPT-6 Astra; Euler stepping stone; ~88 hours; Lean via Astra ~17 hours; millions of dollars of compute.
Mechanism?	Self-similar spaghetti vortex plus oscillatory cancellation so the force stays smooth.
Who else?	Buckmaster–Alpöge (forced Euler/Boussinesq/IPM); Anandkumar et al. (PINN Euler); DeepMind 2025 (unstable singularities).