OpenAI Accused of Scooping Academic Team on Historic Math Problem

OpenAI has published what it describes as a complete proof for the Navier-Stokes existence and smoothness problem, triggering sharp blowback from academic researchers who allege the startup rushed out the solution after learning about their unreleased findings. The dispute, first reported by TechCrunch, centres on one of mathematics' most celebrated challenges: a Millennium Prize problem from the Clay Mathematics Institute carrying a $1 million reward.

Tristan Buckmaster, a mathematics professor at New York University, released three preliminary proofs toward the fluid mechanics problem alongside Levent Alpöge, a mathematician at Anthropic. The pair achieved their results using a combination of OpenAI's Codex and Anthropic's Claude. However, Buckmaster revealed that their milestone was overshadowed by a sudden parallel sprint from OpenAI, which deployed an unreleased frontier model to tackle the same problem using an uncharacteristically similar mathematical route.

OpenAI's counter-release relied on immense computational firepower. The company acknowledged that its effort consumed 300 billion output tokens across a single week, amounting to roughly $22.5 million in compute at current Astra pricing. OpenAI confirmed that its initiative began on September 1, spurred by circulating rumours that solutions to two Millennium Prize problems were imminent.

Allegations of Leaks and Intimidation

According to Buckmaster, word of his team's specific strategy leaked to OpenAI before publication. The researchers had pursued a rare methodology focusing on smooth force options from Charles Fefferman's classic formulation of the problem. Buckmaster argued that this specialised path was not something an AI agent would stumble across in a matter of days without human direction.

When Buckmaster questioned OpenAI about the origins of its project, conversations grew contentious. Buckmaster claimed that researcher Sebastien Bubeck attempted to broker a compromise by requesting that Alpöge, given his role at rival Anthropic, be stripped of credit. Buckmaster stated that when he insisted on going public with the timeline, Bubeck cautioned him by asking, "Why would you ruin your career?" and adding, "If you don't want me to be nice, then I don't have to be nice."

Buckmaster also raised the possibility that his extensive use of Codex might have fed his prompt data back into OpenAI's training pipeline. OpenAI rejected the notion of deliberate data harvesting, writing in a public update that researchers and agents did not access user data, while acknowledging that de-identified telemetry from product usage can influence model behaviour. OpenAI maintained that its final proofs differ substantially in structure from the academic team's work.

The Industrialisation of Pure Math

The clash underscores a growing tension in theoretical science as AI labs wield industrial-scale infrastructure to compete directly with university faculties. When access to raw compute allows a corporation to reproduce and outpace months of bespoke academic research in a few days, the traditional norms of academic priority, attribution, and open collaboration face unprecedented strain.