OpenAI releases 722 AI‑generated math manuscripts on GitHub

What OpenAI released

On Oct 6 2026 OpenAI uploaded a new GitHub repository, openai/math, containing 722 manuscripts that it says were generated by an internal “frontier” model. The papers are organised into 372 families – groups of related results that share a discipline or a common line of reasoning. Each directory holds a PDF, source files, a BibTeX citation block, and, where available, a formal proof written in Lean. The repository is published under an Apache‑2.0 licence.

How the papers were produced

According to the repo’s README, the model was presented with roughly 4,000 open problems. For each successful attempt the team recorded the output, then bundled related outputs into a family. The average compute cost is reported as about three hours of ChatGPT Pro usage per result. Most of the work followed a single‑prompt, single‑agent workflow; two exceptions – a zero‑free region for the Riemann zeta function and a Hodge‑conjecture claim – were edited by humans for readability.

Community reaction

The release has ignited a lively debate. Andrew Sutherland, a mathematician at MIT, warned, “Until the model and the prompts are public, any claim of one‑shot problem solving should be treated as unverified.” Daniel Litt of the University of Toronto argued the opposite, saying, “If we want answers, there’s no reason to keep them secret.” Nobel‑level mathematician Terence Tao has publicly criticised the “insane” speed at which frontier labs are flooding the discipline with AI‑generated results.

The Advisory Group on Mathematics and Artificial Intelligence (AGMAI) – an independent panel of elite mathematicians convened by the Institute for Advanced Study – issued recommendations in late September. The group, after receiving more than 600 responses from the community, urged labs to publish results in scholarly repositories, disclose model name, prompts, compute time and cost, and avoid using releases as marketing. OpenAI says it is “taking the advisory group’s guidelines seriously,” though it has only disclosed the average compute figure and has not released the model itself.

What comes next

OpenAI promises to keep the repository alive, correcting any formalisation errors and adding new Lean proofs over time. It also announced plans for workshops, conferences and summer schools to help mathematicians understand the AI‑generated output. The company maintains that releasing the internal model “as quickly and responsibly as possible” is a priority, but no timeline has been given.

The mathematical community now faces a practical dilemma: how to validate hundreds of new proofs that may be correct in Lean yet lack human‑readable insight, and how to integrate an opaque, proprietary tool into a field built on open, reproducible reasoning. As the papers are digested, the debate over research ethics, credit, and the future of mathematical discovery is only set to deepen.