OpenAI Cracked a $1 Million Math Problem With 10,000 AI Agents, Rival Mathematicians Now Cry Foul

OpenAI said on Tuesday its artificial intelligence cracked a $1 million (£740,000) math problem that has stumped experts since the 1800s, but rival mathematicians accuse the company of rushing to the result after learning they had nearly reached it first.

A Proof Built by 10,000 Agents

The company said an unreleased internal model, which it called far more capable than GPT-6 Astra, directed about 10,000 autonomous AI agents that ran for 88 hours before producing a proof. A separate model then spent another 17 hours formally verifying it in the programming language Lean. The agents exchanged nearly 5 million messages, and OpenAI put the compute bill in the millions of dollars (several million pounds), far above the prize on offer.

The result concerns the Navier-Stokes equations, which describe how fluids move, from blood through arteries to storms across the US. Mathematicians have asked since the 19th century whether the equations can 'blow up', with a tiny patch of fluid accelerating towards infinite speed. OpenAI says its agents found exactly that, a singularity forming in finite time.

Why the Credit Fight Erupted

Hours before OpenAI's announcement, New York University mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge posted their own finite-time blowup proofs for several related fluid problems, built with heavy help from AI models. Both efforts trace back to a strategy pioneered by Diego Córdoba and Luis Martínez-Zoroa.

Buckmaster alleges OpenAI learned of his unpublished work, then pushed a publication plan that would drop Alpöge because he works for a direct competitor. He claims that when he threatened to go public, OpenAI researcher Sébastien Bubeck replied, 'Why would you ruin your career?'

OpenAI Pushes Back

Bubeck rejected that account, calling the claims 'false and inflammatory'. He posted texts he says show he offered a coordinated release and OpenAI's prompts, and called the career remark a frustrated aside he retracted at once.

Sam Altman also denied wrongdoing, saying the two teams' approaches looked different once OpenAI could see the rival work. Altman admitted the company chased the problem only after hearing rumours Anthropic's models had cracked one. 'We were curious if ours could do it too,' he wrote.

A Trust Problem for Everyday Users

The dispute raises a question that touches millions of US professionals who now feed work into ChatGPT, Codex, and rival tools. Startups, law firms, and university labs across the country routinely paste sensitive drafts into these systems.

OpenAI says it did not touch the pair's accounts, yet it conceded it 'cannot rule out' that de-identified data from their use of its products helped train its models. By default, the company says it doesn't train on business customers' inputs, and personal account holders can opt out.

OpenAI won't claim the prize money, framing the work instead as proof of how fast its systems are improving. Its proof tackles the forced version of the equations, listed as statements C and D in the problem's official description, and it still needs independent checking. Six of the seven Clay Millennium Prize problems remain open.

For the millions who now trust a chatbot with unfinished ideas, the worry is simpler: whether today's tool becomes tomorrow's competitor.