OpenAI's $1M Navier-Stokes Claim Comes With an Authorship Fight

OpenAI says an unreleased in-house model has produced a resolution to the Navier–Stokes existence and smoothness problem — one of the Clay Mathematics Institute's seven Millennium Prize Problems, each carrying a $1 million reward since May 24, 2000. If it survives peer review, it would be the first Millennium Prize problem where an AI system did the actual mathematical reasoning rather than just checking a human's proof. Instead, the announcement landed inside a public fight over who saw what, and when — the kind of dispute agentic research workflows are going to keep producing.

What OpenAI says it found

OpenAI's effort reportedly ran the model — on the unreleased line the company calls Astra — as something like 10,000 parallel agent instances over about a week, consuming on the order of 300 billion output tokens across its Millennium Prize attempts, which independent estimates price at $15–22 million against public API rates (TechCrunch, Fortune). The claimed result describes a configuration where a vortex tightens and spins at an ever-increasing rate — a finite-time blowup — while the fluid's total energy stays bounded, which would mean smooth 3D fluid flow can break down under the Navier–Stokes equations. OpenAI says part of the proof has been formalized in the Lean theorem prover. It has not yet been independently verified by the mathematics community.

A parallel track OpenAI didn't start first

NYU's Tristan Buckmaster and Anthropic mathematician Levent Alpöge had spent close to a year working a closely related problem using Claude and OpenAI's own Codex, reaching a breakthrough around August 15. OpenAI's push on Navier-Stokes began September 1 — days after, Buckmaster says, rumors of that breakthrough started circulating. OpenAI's Sébastien Bubeck contacted Buckmaster within days of launching the effort (Simon Willison).

• Buckmaster alleges he was offered a joint release on condition that Alpöge — an Anthropic employee — be dropped from authorship, and that pushback was met with, in his account, "Why would you ruin your career?"
• He also asked whether OpenAI's model had trained on, or had access to, the pair's private Codex sessions from that unpublished work.
• OpenAI's Bubeck disputes the characterization, saying he "came into the discussion following academic norms," and OpenAI states it "did not use their prompts or proofs to prompt our models or direct our agents."

On the data-access question, OpenAI's own statement stops short of a full denial: it says it "cannot rule out that de-identified data derived from their usage of our products helped improve our models." Mathematician Terence Tao, commenting on the episode more broadly, warned that AI labs "strip-mining of open problems for solutions may destroy the ecosystem" that gives researchers a reason to work on them in the first place.

Why builders should care

Strip away the math and the operative fact is structural: usage data from a hosted coding agent can, by the provider's own admission, feed back into that provider's model training — even after de-identification. That's not unique to Codex or Claude, and it's not a hypothetical for anyone doing genuinely novel, IP-sensitive work through an AI pair-programmer. If you're running unpublished research, a proprietary trading strategy, or pre-launch product code through an agentic coding tool, the Navier-Stokes dispute is a concrete reminder to ask, up front, what your provider's data-use and training terms actually say — and to assume behavioral signal can leak even when raw content doesn't.