OpenAI Claims Navier-Stokes Breakthrough as Credit Dispute Erupts
OpenAI has announced a solution to the Navier-Stokes Millennium Prize problem, one of mathematics' longest-standing open questions, using an orchestrated system of up to 10,000 AI agents working together at an estimated cost of approximately $15 million in compute (New Scientist, The New York Times).
The Clay Mathematics Institute established seven Millennium Prize Problems in 2000, each carrying a $1 million reward. Only one had been solved prior to this announcement (Engadget). The Navier-Stokes problem concerns fluid dynamics — the equations that describe the behavior of liquids and gases. OpenAI's own announcement frames them as applying Newton's second law of motion (F=ma) to fluid flow (OpenAI). The specific result, presented in a PDF hosted on OpenAI's CDN, describes a solution that starts from rest and develops unbounded velocity in finite time while maintaining uniformly bounded kinetic energy (OpenAI PDF). Mathematicians refer to such hypothetical solutions, where fluid speed becomes infinite, as "blowups" (Scientific American).
The achievement is shadowed by a dispute over priority, credit, and data access. Mathematicians Tristan Buckmaster of New York University and Levent Alpöge, a researcher employed at Anthropic but working in a personal capacity, had been developing their own Navier-Stokes solution using multiple LLMs: Anthropic's Claude, OpenAI's Codex tool, and OpenAI's Astra frontier model (Engadget). According to Buckmaster, OpenAI may have intensified its Navier-Stokes efforts after learning that a team including rival Anthropic was close to a solution. Buckmaster questioned whether OpenAI had leveraged the work he and Alpöge had conducted through Codex, the OpenAI coding tool into which they had input all their project drafts.
Buckmaster said he asked whether the model had been trained on or had access to the Codex sessions containing their drafts. He was told the model did not look up user data. When he asked again specifically about training, he did not receive an answer (Engadget).
The dispute over credit is equally pointed. Buckmaster said he spoke with Sebastien Bubeck, a mathematician and AI researcher at OpenAI, and another unnamed party, and was offered two options. The first was to publish a partial development, with OpenAI to publish its own Navier-Stokes solution the following day. The second was to write a solo paper crediting the use of an OpenAI LLM in the solution, without including Alpöge's name. Buckmaster declined both proposals (Engadget).
The broader context here involves unresolved questions about how AI labs handle user data generated through their own tools, and what obligations they have when internal research teams may benefit from that data. The question of whether Codex session inputs were used to train or inform OpenAI's models goes to the heart of trust in AI-assisted research workflows. If researchers cannot be confident that their unpublished drafts, proofs, and methodologies entered into a tool are not subsequently absorbed into the lab's own capabilities, the calculus of using frontier models for sensitive intellectual work changes substantially.
Terence Tao, one of the most prominent mathematicians working today, offered a cautionary note well before the announcement. Tao commented that an unreadable but formally verified ChatGPT proof of the Navier-Stokes problem would be "a disaster for the field" (Mathstodon). His concern cuts to a different issue than the credit dispute: whether a solution that no human can meaningfully understand or verify by hand advances mathematics, or merely satisfies a formal criterion. The tension between machine-generated proofs and human mathematical understanding has been building for years, and a Millennium Prize solution produced by 10,000 AI agents places that tension at the center of the field.
Separately, Scientific American reported that Alpöge used his AI-assisted method to prove that the Euler equations, the frictionless counterparts to the Navier-Stokes equations, do in fact blow up (Scientific American). That result, if it holds, is a related but distinct advance.
The parallel research threads, the competitive dynamics between OpenAI and Anthropic, and the unresolved data-access questions mean the Navier-Stokes announcement will likely be debated on multiple fronts for some time. The Clay Mathematics Institute has not yet publicly weighed in on whether OpenAI's solution meets its criteria for the $1 million prize. The technical result will require rigorous verification, and the circumstances of its discovery may shape how the mathematics community, and AI-assisted researchers in particular, approach collaboration with frontier model providers.
