October 11, 2026
An unreleased OpenAI model produced math results: 372 families are on GitHub
OpenAI published 372 families of mathematical manuscripts on GitHub on Oct 6, 2026. The model that produced them has not been released: no name, no API, no prices.
Out of roughly 4,000 problems posed to the model, 372 families of manuscripts grew. The README does not state the share of solved problems.
What is open. The openai/math repository: PDF preprints, source files, the CONTENTS.md manuscript map, an overview PDF, and a Lean library with a catalog of formalizations. License: Apache-2.0. Start with the overview PDF and CONTENTS.md: the model itself cannot be run. One result took about three hours of “ChatGPT Pro thinking,” per OpenAI’s wording.
What is verified. OpenAI itself warns that some informal results may contain errors. Fixes are promised, with no timeline. A result without a Lean proof should not be considered verified.
Training of the internal model began on Aug 28, 2026; evaluations started on Sep 1. OpenAI says it is working on a responsible release of the model.
OpenAI has not given a timeline for fixes or for the model’s release.