Title: OpenAI Drops 722 Math Manuscripts at Once — and the Math World Is Reeling
October 7, 2026
On October 6, OpenAI published 722 mathematical manuscripts in a public GitHub repository, grouped into 372 families of related results. The results were produced by an unreleased internal AI model: during the evaluation the model was given about 4,000 math problems, with each result using, on average, computing equivalent to roughly three hours of ChatGPT Pro "thinking." Some proofs come with versions written in Lean — a language that lets a computer check a proof step by step — though not all do yet. The New York Times called the release a move that "further roils" an already turbulent field.
The most eye-catching claims include a proof of a "quasi-Riemann Hypothesis" — establishing a zero-free region for Dirichlet L-functions with real part greater than 7/8 and eliminating Landau-Siegel zeros — reportedly verified with Lean. Other highlights: an unconditional proof of NP-hardness for the semidefinite programming approximation threshold, the rational Hodge conjecture for CM abelian varieties, and the symmetric and asymmetric Mahler conjectures in convex geometry. Daniel Litt of the University of Toronto said that if mathematicians want the answers, he sees no reason to ask the company "to keep them secret from us." Rutgers' Alex Kontorovich called one finding "insane," saying a human achieving it would "earn a Fields Medal instantly."
The controversy is about transparency. On September 29, the Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, published recommendations: AI labs should stop testing advanced math problems on proprietary models; for results the prompters don't yet understand, they should name the model and publish the prompts, a reasoning summary, and the computing cost; and math results shouldn't be used as model marketing. OpenAI said it takes the guidelines seriously but isn't bound by them: it shared average compute statistics and reasoning summaries for 10 results, but no prompts and no model name. The Wall Street Journal reported that in the days before the release, mathematicians scrambled to publish their own AI-assisted work before getting scooped.
An audit posted to arXiv reviewed the 10 results OpenAI announced on August 1: it found no confirmed substantive mathematical error in a principal result within the reviews examined, though review depth varied — and it doesn't cover the October manuscripts. The key caveat is the field's consensus: a proof published by an AI company doesn't become established mathematics until it survives scrutiny by expert mathematicians. OpenAI itself says it does not intend to claim the Millennium Prize. The math world's task now is to verify the 722 manuscripts one by one. If they hold up, it's a milestone; if they don't, it's a cautionary tale. Either way, the way mathematics gets done has already changed.
Sources: The New York Times, The Wall Street Journal, The Verge, Scientific American, Interesting Engineering, AI Stock Wire, CNBC-TV18