Mathematicians accuse OpenAI of stealing Navier-Stokes solution

OpenAI claimed a historic mathematical breakthrough, but mathematicians call it theft.

On September 8, 2026, OpenAI researchers Metab Swani and Mark Selke outlined on The a16z Show how frontier models were cracking decades-old math problems. Models like GPT-6 Astra demonstrated autonomous reasoning by backtracking through complex geometric bounds and pruning dead-end proofs without carrying human mental clutter. Swani and Selke argued that AI was transforming pure mathematics, shifting human effort from tedious derivations to high-level curation.

That vision of seamless human-machine collaboration collapsed days later over the Millennium Prize Navier-Stokes problem. OpenAI announced that an autonomous swarm of 100 agents running on GPT-6 Astra solved the existence and smoothness equations in under four days. The company burned through over $1 million in compute, deploying interconnected agents to generate and verify solution paths.

The timing immediately raised alarms across the academic community. On This Week in Startups on September 9, 2026, NYU mathematician Tristan Buckmaster alleged that OpenAI front-ran his breakthrough. Buckmaster and Anthropic researcher Levant Alpoge had spent months drafting proofs for the 90-year-old equation, regularly running their private code through OpenAI's Codex interface.

Buckmaster revealed that OpenAI launched its internal Navier-Stokes project only after he and Alpoge ran their work through Codex in September. As the human researchers closed in on publication, OpenAI deployed its computational brute force along the exact same unconventional path. Buckmaster further claimed that OpenAI executives pressured him to drop Alpoge from the project due to corporate rivalry with Anthropic.

OpenAI executive Sebastian Bubeck rejected allegations of intellectual property theft on This Week in Startups, maintaining the company acted on broader industry rumors. However, as reporting on Breaking Points on September 10 confirmed, OpenAI could not rule out that user prompt sessions were ingested into model training pipelines. When Buckmaster pressed executives about data privacy, a company representative warned him against endangering his academic career.

By September 11, 2026, The Intelligence reported that OpenAI waived the $1 million Millennium Prize. Science correspondent Sam Weigley noted that OpenAI's published paper omitted the rigorous structural proofs required by traditional peer review. Instead of presenting a clean conceptual framework, the lab delivered agent output log traces that left mathematicians questioning whether the swarm achieved genuine insight or simply ingested Buckmaster's private work.

The fallout extends far beyond a single equation. On This Week in Startups, Canvas Ventures co-founder Rebecca Lynn cautioned researchers that commercial frontier models inevitably absorb user inputs. Agent Fund general partner Yohi Nakajima observed that autonomous indexing agents are fundamentally dismantling academic privacy, turning every private query into potential training data for tech monopolies.

AI tools promise to accelerate scientific discovery, but they risk eating their own sources. When researchers cannot trust commercial tools with early hypotheses, open scientific inquiry shuts down.