Two OpenAI researchers called an NYU mathematician Tristan Buckmaster on Sept. 6 to tell him a machine just finished the proof he had spent a year solving, by the same narrow route almost no one else was trying.

The Culture and Clarity newsletter reported that, by the end of the second call, according to the mathematician, he had been offered a share of the credit if he dropped his co-author. When he said he would go public, he was asked why he would want to ruin his career.

Two days later, both sides went public, and the fight over $1 million math prize turned into a case study every executive using AI tools should read.

What OpenAI Says It Solved

The Navier-Stokes equations describe how water, air and smoke move. Engineers use them daily but no one had proven whether they always behave or can produce infinite speed at a simple point. This is what mathematicians call a blowup. The Clay Mathematics Institute has since 2000 offered $1 million for an answer.

Buckmaster and Levent Alpoge, a mathematician employed by Anthropic, worked on a way to solve the blowup proof for most of a year as a personal project with no institutional backing. Buckmaster paid for the AI tools out of his own research funds, including what he calls a large bill from OpenAI. The mathematicians stored every draft in OpenAI’s Codex coding agent.

Hearing rumors of OpenAI’s progress, Buckmaster emailed a senior OpenAI mathematician to hear what was happening.

OpenAI said it had already pointed an unreleased model at the problem on Sept. 1, and went on to run 10,000 agents for 88 hours, and had a full proof on Sept. 5. In a series of phonecalls the next day, Buckmaster asked whether the model had been trained on his Codex sessions. He said the question went unanswered.

Buckmaster is careful about what he is claiming. He has not seen OpenAI’s proof, does not know whether his data was used and writes that he is not accusing anyone of anything, only stating what he was told and what was proposed.

OpenAI denied using Buckmaster and Alopge’s work, said it will not claim the prize money and puts its compute cost in the millions of dollars. However, OpenAI said it cannot rule out that the de-identified data from its usage helped improve the models.

Sebastian Bubeck, who leads OpenAI ’s math team, confirmed he said it would be simpler if Alpoge did not work at a competitor, and confirmed that the career remark was a bad choice of words. He said the offer was for Buckmaster to author a write up of OpenAI’s proof, not to remove Alpoge from his own paper. OpenAI CEO Sam Altman backed that version.

The Clay Institute still lists the problem as unsolved and has promised a very detailed review of the whole situation.

Question 1 For OpenAI Customers: What Are You Putting Into The Agent

Buckmaster’s suspicion is unproven as of today. It still matters because he was a paying customer asking exactly what your general counsel will ask after your product team pastes six months of roadmap into a coding agent. Questions like, “does the vendor train on it, and who at the vendor can see it?”

OpenAI’s statement that no specific user data was accessed answers the second question but not the first. Data transparency is important. If your AI contracts do not spell out training rights, retention and employee access in language that is enforceable, you need to look at fixing that for the next renewal.

Question 2: Who Gets Credit When OpenAI’s Model Does The Work?

Most companies have no policy for attributing work to an AI system produced, no rule for what happens when an employees’s agent output resembles a partner’s and no plan for a credit dispute with a vendor. Two of the best resourced organizations in the field could not agree on authorship of a single document. Write your policy now, while nothing is at stake.

Question 3: Is Rivalry Driving Your AI Spend The Way It Drove OpenAI’s?

OpenAI’s own account posted that the company aimed its best model at this problem because of a rumor about a competitor. That is a multimillion-dollar week of work justified by fear of losing a headline. Boards should ask whether their own AI budgets are set by what a rival announced last quarter or by problems the company needs solved. Rivalry is a fine motivator and a poor strategy.

Question 4: Are Your Buying Answers From OpenAI or Building Capability?

Buckmaster said the results are not the important part. What matters is that a mathematician and a model can now do a year’s worth of work in a month, which he called a Deep Blue-Kasparov moment for how the field trains students and assigns credit.

He was candid about the cost. He wrote that the machine generated proof was the most horrendous he had ever read, his team had been working around the clock to understand it, and he called his own rushed Euler note “AI Slop.”

Terrence Tao, one of the most respected mathematicians, warned this week in Yahoo Tech that harvesting open problems for answers may destroy the ecosystem that produces the next generation of talent.

Leaders who have AI produce answers without building understanding in their teams will win the quarter but lose the decade.

Whether or not OpenAI’s proof survives review, the contract you signed with your AI vendor is already in force, and it is worth reading again this week.