The largest AI-generated mathematical publication ever
OpenAI published on October 6, 2026 the largest set of mathematical results ever produced by an artificial intelligence system: 722 manuscripts organized into 372 "result families," containing 377 new proofs or mathematical contributions. The work was generated by an unreleased frontier model, previously referred to internally as "Astra," and published on the company's GitHub with clear protocols for review and citation.
The scale is overwhelming. The manuscripts cover advances in approximately 4,000 open problems in areas including number theory, geometry, graph theory, complexity theory, and cryptography. Among the highlights are contributions toward the Birch-Swinnerton-Dyer leading-term formula for elliptic curves over the rational numbers — under specific conditions — a zero-free region for the quasi-Riemann hypothesis beyond Re(s) > 7/8, and a new upper bound of 9/4 on the matrix multiplication exponent.
The company shared the manuscripts under a set of protocols designed to integrate with existing academic practice, including revision procedures and citation guidelines.
The ten landmark proofs
On its own blog, OpenAI selected ten results as representatives of this massive collection. Each one resolves or significantly advances a long-standing problem in its community: from high-dimensional sphere packing to binary and spherical codes, non-sofic groups, Connes's rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, the closest vector problem (related to post-quantum cryptography), Ehrhart's volume conjecture, multicolor Ramsey numbers, and extremal conjectures in graph theory.
What makes this collection different from the company's previous publications — such as the disproof of Erdős's unit-distance conjecture in May — is the complete formalization in Lean. Each argument was formally certified in Lean, a computer-assisted proof system that allows logical verification step by step. This is not merely a technical detail: it represents a change in the validation standard when results are produced by AI.
Compute costs and production scale
According to OpenAI, each individual result used, on average, compute equivalent to about three hours of ChatGPT Pro-level reasoning — which, at Sol API rates, would amount to approximately $2,000 for solving the full set of ten selected proofs. The company's human authors prepared the manuscripts from the arguments generated by the model and then formalized each proof in Lean to produce verifiable certificates.
This places AI mathematical production in a completely new category: it is no longer about assisted insight, but a large-scale production pipeline where AI generates, prepares, and formalizes hundreds of mathematical contributions in days, not years. For comparison, the mathematicians who solved the Navier-Stokes problem (announced by OpenAI in August) spent months of work.
The mathematical community's response
The publication did not go unnoticed. Mathematicians expressed concern both about the speed of the advance and the ethical and academic implications of automated proof production. In response, OpenAI established the "Advisory Group on Mathematics and Artificial Intelligence" hosted at the Institute for Advanced Study — an independent committee with members including Timothy Gowers (Collège de France and Cambridge), Edward Witten (IAS), Martin Hairer (EPFL and Imperial College London), and Melanie Matchett Wood (Harvard), among others.
The group operates independently: members are unpaid volunteers who can recommend membership changes, offer advice the company has not solicited, comment publicly on the effects of AI on mathematics, and publish their own conclusions. The company described the initiative as a first step and acknowledged that difficult questions remain about how AI can deepen mathematical understanding and how its benefits should reach the broader community.
What to watch
What sets OpenAI apart from all other AI agents that have attempted mathematics — and what makes this release particularly significant — is not just volume. It is the fact that the results are available under protocols that allow the mathematical community to review, contest, cite, and build on them in a structured way. The GitHub repository contains both the manuscripts and summaries of the model's reasoning process, compute estimates, and statistics on attempted problems.
The question is whether 21st-century mathematics will be defined by human discovery assisted by AI, fully autonomous discovery, or something we do not yet have a name for — but which is already happening, in 722 manuscripts that have just been released into the world.
Sources: OpenAI, The News International, The Verge
✓ Independent sources cross-checked and verified before publishing
