How OpenAI's AI Mistranslated Its Own Navier-Stokes Math Proof Into Code

What OpenAI announced

OpenAI said on September 8, 2026, that an internal multi-agent system had produced a proof addressing the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems carrying a $1 million award from the Clay Mathematics Institute, according to the AI-tracking site implicator.ai. The company reported that roughly 10,000 agents worked for 88 hours starting August 28, generating 2.7 million agent messages and about 130 billion output tokens, before a separate model, described as GPT-6 Astra, spent 17 more hours formally checking the argument in the proof language Lean.

The result was a natural-language manuscript titled "Finite time blowup for Navier–Stokes," accompanied by a Lean project meant to let computers check every logical step, according to implicator.ai and the arXiv paper that later analyzed the work.

The translation-error findings

A paper posted on the preprint server arXiv, titled "Navier-Stokes Lost in Translation" and written by mathematicians Alexander Bastounis, Fabian Circelli and Anders C. Hansen, examined whether OpenAI's Lean code actually matches the mathematical argument in the natural-language paper. The authors distinguish between a Lean proof that merely compiles and one that is semantically faithful to the original reasoning, arguing the second kind is far harder to guarantee than the first.

The paper lists concrete mismatches. In one case, the natural-language paper's Lemma 8.6 requires bounding derivatives using four additional orders of input, written m+4, but the matching Lean theorem instead requires five additional orders, m+5 — a stronger condition than the paper states. In another, equation 10.19 of the manuscript bounds a pressure-flux term using only one variable, called B_R, while the Lean version of the same bound also depends on a second variable, A_R, derived from a different mathematical technique.

The authors conclude that a Lean file compiling successfully only shows that some correct proof exists for the stated theorem, not that it is the same proof described in words. They write that OpenAI's Lean code often imposes stronger conditions on the bounds of various functions than the corresponding results in the natural-language paper, and they recommend that OpenAI's Navier-Stokes and related Euler-equation formalizations undergo the same peer review traditional proofs receive before being trusted.

A separate dispute: which problem was solved?

Independently of the translation question, several mathematicians disputed whether OpenAI solved the version of Navier-Stokes most researchers care about. The Clay Institute's official 2000 problem statement, written by Charles Fefferman, includes an option labelled C that allows a smooth external forcing term, while most mathematicians have focused on the unforced case, Scientific American reported.

OpenAI's proof targets option C. Luis Silvestre of the University of Chicago and colleagues posted a paper arguing that OpenAI's blow-up mechanism cannot work once the external force is removed, Scientific American reported.

Part of a wider pattern of scrutiny

Scrutiny of OpenAI's math claims is not new. In October 2025, then-OpenAI vice-president Kevin Weil said GPT-5 had solved ten previously unsolved Erdős problems and made progress on eleven more; mathematicians including Thomas Bloom said the results already existed in published literature, and Weil withdrew the post, TechCrunch reported. When OpenAI announced in May 2026 that a reasoning model had disproved an Erdős geometry conjecture dating to 1946, it published supporting statements from mathematicians Noga Alon, Melanie Wood and Thomas Bloom — the same Bloom who had flagged the earlier error — alongside the claim, TechCrunch reported.

On October 7, 2026, OpenAI published 372 AI-generated mathematical results on GitHub, each meant to solve an open problem or make substantial progress toward one, most produced from a single prompt to an internal model at a cost of roughly three hours of ChatGPT Pro computing time each, the-decoder reported. Twenty-five Fields Medal winners signed an open letter warning that mass-producing true mathematical statements risked harming the field rather than advancing it, according to the same report, which noted most results had not gone through standard peer review.

Where the claim stands now

The Clay Mathematics Institute still listed Navier-Stokes among its unsolved Millennium Prize Problems, implicator.ai reported. Institute president Martin Bridson said the announcement was exciting but that any formal evaluation would be unhurried and rigorous.