9 October 2026
OpenAI releases 700-plus machine-generated math manuscripts from unreleased model
First reported
Ars Technica ran this on , a day before the other 2 sources picked it up.
- OpenAI released about 720 mathematical manuscripts, grouped into 372 topic families, all produced by an internal model it has not made public.
- The papers were generated with an average of roughly three hours of heavy reasoning time each, according to the newsletters.
- Some results are checked by Lean, a computer system for verifying proofs, but not all of them are.
Our read
Two of the 30 tracked newsletters covered it, and the two differed: Last Week in AI reported the AGMAI board's disclosure request that the company apparently did not follow, while Exponential View focused on the release's scale. This is the third story in our archive on the same release, following items dated 2026-10-08 and 2026-10-10.
Earlier on this story
Written by Newskeryx from which newsletters covered this story and from our own archive.
Where they differ
Exponential Viewinstead highlights the scale of the release and a claim that it accounts for most major math discoveries of the past three years.
Last Week in AIis the one that reports the AGMAI board's disclosure request and that the company apparently did not follow it.
What each one covered
The OpenAI math release, its scale and families, Lean verification, and the split between machine and human mathematics
The 719 manuscripts, Lean formalizations, compute and usage stats, advisory board recommendations, and criticism over disclosure
Reported by Ars Technica