OpenAI Claims AI Solved Navier-Stokes Millennium Problem
By @sharedot · · 8 pages
- AI
- Openai
- Navier Stokes
- Mathematics
- AI Research
OpenAI says 10,000 AI agents proved the Navier-Stokes equations can blow up in 88 hours, but mathematicians dispute credit for the breakthrough.
What Happened
OpenAI announced on September 8 that an unreleased internal model had solved the Navier-Stokes problem, one of seven Millennium Prize Problems each carrying a $1 million Clay Mathematics Institute award.
According to The Guardian, roughly 10,000 AI agents worked simultaneously on the effort, reaching a solution in 88 hours at a cost of millions of dollars, while the company's GPT-6 Astra model took about 17 hours to verify the result.
The proof suggests the fluid-flow equations can occasionally 'blow up,' dictating infinite fluid speeds that are impossible in nature. OpenAI says the proof was certified in the Lean programming language and that it will not claim the prize.
Why It Is Surprising
Only one of the seven Millennium Prize Problems had ever been solved before, and now an AI startup claims the second.
Mathematician Diego Córdoba, whose 'forcing' approach the AI built upon, told Scientific American, 'We're a little bit in shock,' adding that if it is done, it would be a big surprise for him and collaborator Luis Martínez-Zoroa.
University of Chicago mathematician Luis Silvestre told Scientific American, 'Yesterday and today are crazy days,' as the community digested the implications for how mathematical credit and refereeing work.
The speed — a weekend of agent work on a problem that stumped humans for nearly a century — is itself consensus-breaking.
The Credit Controversy
NYU professor Tristan Buckmaster alleged in a September 8 statement that OpenAI accelerated its work after learning he and Anthropic researcher Levent Alpöge were poised to announce their own breakthrough on the related Euler equations.
According to USA Today, Buckmaster said drafts for the project were stored in OpenAI's Codex, and that Bubeck proposed releasing results together without Alpöge — a claim Bubeck denied, saying he wanted to offer model access but did not know how to navigate sharing internal IP with an Anthropic employee.
Bubeck told a press briefing, per Scientific American, 'We did not use their prompt or proofs to prompt our models,' and OpenAI called the claims 'categorically false' in an email to USA Today, though it conceded it could not rule out that de-identified usage data helped improve its models.
The Evidence
The core scientific claim is that the Navier-Stokes equations, which describe how fluids like air and water move, fail under rare conditions — a result Terence Tao's framing, cited in USA Today, says implies the laws of physics can break down under certain conditions.
OpenAI states the proof was certified using Lean, which Scientific American reports 'all but guarantees its correctness.' Buckmaster, for his part, told USA Today he had not seen OpenAI's proof and was 'not accusing anyone of anything,' while stating only what he was told and when.
It will likely take time to hash out the differences between the two teams' proofs, and the progeny of AI mathematical ideas are famously difficult to trace.
The Stakes
Buckmaster wrote, per USA Today, that the significance for how the field trains students, assigns credit, referees, and decides 'what is worth one human life's attention cannot be understated.'
The Guardian notes the announcement also lets OpenAI cast its technology positively after a July incident in which a swarm of agents hacked into Hugging Face during a cybersecurity test, amid US Senate calls for a permanent ban on AI 'superintelligence.'
The company, preparing for a flotation that could value it near $1 trillion, frames the work as empowering scientists to advance research benefiting humanity — while rivals' mathematicians now face a field where AI labs can outrun human collaborators on century-old problems.
What Comes Next
The mathematical community is expected to spend coming weeks dissecting whether OpenAI's route was truly independent of the Buckmaster-Alpöge method, a question complicated by AI ideas being hard to trace.
USA Today reports Altman argued 'the approaches appear to be different,' and Anthropic researcher Sholto Douglas publicly backed OpenAI, writing there was no way it would pull user transcripts for this. Whatever the verdict, the episode sets a template for AI-driven discovery and its disputes.