Story
September 9, 2026
OpenAI’s Navier–Stokes Triumph Is Colliding With a Trust Crisis
OpenAI portrays its result as an independently produced leap in AI mathematics; Tristan Buckmaster and Levent Alpöge see unanswered questions about data, credit and the risks of trusting private labs with unfinished research. The dispute is also a warning that scientific prestige may increasingly hinge on compute controlled by a handful of companies.
On September 1, OpenAI says, researchers heard rumors that two Millennium Prize problems had been resolved and set a new internal model loose on major open questions. Its multi-agent system—roughly 10,000 agents with code tools and a cached internet—reached a Navier–Stokes result about 88 hours later, the company says.1
The claimed proof says a smooth three-dimensional fluid can develop a finite-time singularity: a breakdown in the equations used to model phenomena from weather to blood flow. OpenAI released a written proof and a Lean formalization, calling it a resolution of the existence-and-smoothness problem, one of mathematics’ seven $1 million Millennium Prize Problems.1 Sam Altman celebrated it as “one of the most amazing moments for me in OpenAI history.”
2
But the chronology quickly became the controversy. On the preceding day, NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge published work on a related forced Euler problem after nearly a year of AI-assisted collaboration. Buckmaster said their route resembled OpenAI’s in a way that was “not the direction one arrives at in a few days by giving a model the problem statement.”3 He asked whether the company’s model had trained on, or accessed, their Codex sessions containing project drafts—while stressing that he had not seen OpenAI’s proof and did not know whether their data was used.3
OpenAI denies that researchers or agents saw the pair’s work before its public release, and says the Euler results differ: its agents solved an unforced version while Buckmaster and Alpöge worked on a forced one. Yet its statement leaves a narrow but consequential caveat: it “cannot rule out” that de-identified usage data helped improve its models.1
That qualification has amplified calls for scrutiny. Yann LeCun boosted a demand to hear “alternative points of view on how it happened.”
4 The stakes extend beyond priority: critics fear a future in which researchers using commercial AI tools can be outpaced—and potentially competed against—by the firms that operate them, armed with private models and millions of dollars in compute.5