OpenAI Dropped 722 Math Papers From an Unreleased Model, and Mathematicians Want Receipts

OpenAI published a batch of mathematics papers on the evening of Tuesday, October 6, 2026, and by the next morning almost nobody had read them. The company's own repository counts 722 manuscripts grouped into 372 result families, all "produced by an internal OpenAI model" that has not been released. 12

The model was posed roughly 4,000 problems, and the average published result used the equivalent of about three hours of ChatGPT Pro thinking compute. An OpenAI spokesperson told Scientific American that almost every result came from a single prompt handed to a single AI agent — a different shape of job from the company's earlier Navier–Stokes work, which was produced by a 10,000-strong agentic swarm that cost millions. 13

Many of the proofs ship with Lean formalizations, which a computer can check, and OpenAI says not all of them do. The README adds, in the company's own words, that "some of the unformalized results could have issues." What was disclosed alongside the papers is ten summaries of the model's reasoning, an average compute figure and statistics on how many problems were attempted — no prompts, no per-result compute and no model. 12

Among the results claimed are a solution to the four-dimensional Kakeya conjecture, a zero-free region for the Riemann zeta function past eleven twelfths, the Hodge conjecture for CM abelian varieties, a Birch–Swinnerton-Dyer leading-term formula for some elliptic curves, and improvements to established computer algorithms. Mathematicians told Scientific American that working out whether these contain novel ideas or are mostly mash-ups of existing techniques will take months. The counts circulating in the coverage do not even agree: 722 manuscripts, 372 families, 377 results. 345

The independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, which OpenAI says it consulted, had already asked labs to publish the model name, the prompts, a summarized chain of thought, the time taken and the estimated compute cost for each result, and to stop treating releases as marketing vehicles. OpenAI says it is not bound by those recommendations. MIT's Andrew Sutherland told Scientific American: "Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified. We should ask for receipts." 36

Terence Tao put the moment in his own frame: "Math 1.0" rewarded whoever solved a problem first, and that goal "has been optimized to the point of unsustainability"; "Math 2.0" will have to value exposition, community and new directions instead. OpenAI acknowledges that many of the newly released results are not yet understood by its own mathematicians. Nobody has checked the bulk of the pile. 37

"Show Your Work" is an original rap song and music video built from those facts.