TL;DR
OpenAI's multi-agent AI system claims to have solved the Navier-Stokes Millennium Prize Problem in 88 hours, a result that would be the most significant AI-assisted mathematical breakthrough in history, though it arrives wrapped in a priority dispute and still awaits peer review.
What happened
- On September 8, 2026, OpenAI published a claimed proof for the Navier-Stokes existence and smoothness problem, one of seven Clay Millennium Prize Problems each worth $1 million.
- The proof was produced by a next-generation model more capable than GPT-6 Astra, running approximately 10,000 coordinated AI agents simultaneously.
- Agents exchanged 2.7 million messages and generated roughly 130 billion output tokens on Navier-Stokes alone; the full math experiment totaled 4.9 million messages and 300 billion tokens.
- The result: a constructed singularity, a vortex that twists inward and accelerates without bound under smooth external forcing, satisfying versions C and D of the official problem statement. It does not prove singularities arise in ordinary unforced fluids.
- GPT-6 Astra took 17 additional hours to formalize and verify the proof in Lean. OpenAI says it will not claim the prize money.
Why it matters
- Navier-Stokes underpins fluid dynamics across aviation, weather forecasting, and medicine. A confirmed singularity result reshapes foundational assumptions in applied mathematics and physics.
- The compute scale, described by Nature as unprecedented for a single mathematical study, signals that frontier AI can now mount sustained, massively parallel attacks on century-old hard problems.
- A priority dispute clouds the result: NYU mathematician Tristan Buckmaster and Anthropic-affiliated Harvard mathematician Levent Alpöge published related singularity results on September 7, one day before OpenAI. Buckmaster alleges OpenAI accelerated after learning of their unpublished work stored in Codex, and claims OpenAI's research head suggested excluding Alpöge from authorship due to his Anthropic affiliation.
- OpenAI acknowledged it cannot rule out that researcher interactions with its products influenced model training, a disclosure that raises serious questions about data boundaries between users and model developers.
- Under Clay Institute rules, the proof must be published in a qualified journal and gain broad community recognition over at least two years before official evaluation, meaning the $1 million prize is years away at minimum.
What to watch next
- Peer review and journal submission: whether the proof survives scrutiny from the global mathematics community is the single most important signal. Watch for responses from leading fluid dynamics and PDE mathematicians.
- Resolution of the priority dispute: Buckmaster's allegations about Codex data access and authorship pressure could trigger formal investigations with implications for how AI labs handle researcher data.
- Replication attempts on remaining Millennium Problems: OpenAI's agents also produced a result on Euler equation regularity. Whether the same system is now directed at the other five unsolved problems (Riemann Hypothesis, P vs NP, etc.) would confirm this is a repeatable capability, not a one-off.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.