OpenAI’s claim that an unreleased model solved the Navier-Stokes existence and smoothness problem has opened a second contest alongside the mathematics itself: who deserves credit, what information influenced the effort and whether researchers can trust AI companies with their work in progress.
The problem carries a $1 million Clay Mathematics Institute prize and is one of the Millennium Prize Problems. According to TechCrunch, OpenAI ran a week-long effort that produced roughly 300 billion output tokens. The source estimates that volume would cost $22.5 million at current Astra rates, although that comparison does not establish OpenAI’s internal cost.
Parallel Proofs Arrived Within Days
New York University mathematician Tristan Buckmaster and colleague Levent Alpöge announced three proofs on Tuesday. Their work included a preliminary result on a major theoretical problem and used both OpenAI’s Codex and Anthropic’s Claude. OpenAI then presented what it described as a complete proof of the Navier-Stokes problem, created with a next-generation model that has not been publicly released.
The timing matters because OpenAI acknowledged that its latest effort began on September 1 after hearing rumors of progress and through conversations about the researchers’ work. Buckmaster says information about their direction reached OpenAI and that its team pursued the same uncommon route. His account presents the lab’s result as a rush triggered by knowledge of work already underway.
OpenAI says the resulting proofs are different and argues that it did not use the researchers’ prompts or proof. The available source does not independently settle either the mathematical priority or the pathway by which each team reached its result. It establishes that multiple efforts converged quickly and that the participants sharply disagree about the significance of the overlap.
Credit Became a Career-Level Dispute
Buckmaster also accused OpenAI researcher Sébastien Bubeck of asking him to remove Alpöge from credit and of making a threat about Alpöge’s career. Those are Buckmaster’s allegations, not independently verified findings. OpenAI disputes his characterization of events.
The disagreement exposes the stakes around attribution. A proof attached to a famous open problem could shape academic reputations, while a successful AI-generated result could become a major claim for the company that built the model. When those incentives collide, informal conversations and partial disclosures can quickly become evidence in a dispute over priority.
The source also raises a separate concern about researchers using commercial AI tools. Buckmaster and Alpöge worked with Codex and Claude while developing their results. OpenAI says it did not access their specific user data for its proof effort. It also says it cannot rule out the possibility that de-identified product usage had previously helped improve models. That distinction leaves the central concern unresolved: there is no evidence in the article that their private work was directly retrieved, but broad model improvement can make provenance difficult to trace.
Verification Must Be Separated From Publicity
A claimed solution to a Millennium Prize Problem requires scrutiny beyond a company announcement. The article reports OpenAI’s result as a claim and does not establish that the mathematical community or the Clay Mathematics Institute has accepted it. The cost, scale and speed of the computation may be notable, but none substitutes for checking every logical step.
The same discipline should apply to the allegations. OpenAI’s admission that rumors and conversations inspired its effort supports part of the timeline, while Buckmaster’s stronger claims about copied direction, credit and pressure remain contested. Treating either side’s account as settled would go beyond the source.
The episode demonstrates how AI changes the mechanics of research competition. Thousands of parallel model runs can compress exploration into days, but they also make rules about confidentiality, attribution and model-training data more consequential. Before the mathematics is resolved, the dispute has already shown that research institutions and AI labs need clearer expectations for handling unpublished ideas and credit when humans and models work on the same frontier.
