AI news · October 8, 2026
OpenAI publishes formal math progress without releasing the underlying model
OpenAI's public math repository lists 722 manuscripts in 372 families and says each result used about three hours of ChatGPT Pro thinking compute on average. The model itself is not released.
- manuscripts in the public catalogue
- 722
- result families
- 372
- average ChatGPT Pro thinking compute per result
- 3 hours
OpenAI published a collection of mathematical results on October 6 and linked a public GitHub repository for the manuscripts and supporting proof artifacts. The repository lists 722 manuscripts organized into 372 families, with revisions preserved and Lean formalizations available for some results. OpenAI says most results came from an unreleased internal model and used about three hours of ChatGPT Pro thinking compute per result on average. The evaluation posed about 4,000 problems before results were grouped and selected.
This is a research release, not a new model that builders can call. The practical value is in the review material: the repository exposes examples of model-produced mathematics, formal proof work, citations, and known gaps. Anyone using AI for technical research should read the result status and formalization notes before repeating a claim. The collection is also a reminder that compute, verification, and human selection remain part of the final result.
What you can do with it
Use the public repository as a review exercise rather than a production solver. Pick one manuscript, inspect its status and supporting files, and check whether a Lean formalization exists. Record what is machine-produced, what is verified, and what still needs a human mathematical review.
Our take
Publishing the artifacts is more valuable than a polished claim because it gives researchers material to inspect and correct. It does not prove that the unreleased model can replace mathematicians, and the repository itself warns that some work is not formalized. Treat it as an evidence set, not an answer engine.
Links Research post Public repository HN discussion
Source: OpenAI ↗ — Made With Models writes the brief; the reporting is theirs.