The typical result in OpenAI’s new batch of mathematics took about three hours of Pro-tier thinking on ChatGPT, by the company’s own estimate. That is compute priced as a consumer subscription, not as hardware, so it is easy to picture and hard to audit.
The results come from an internal model that nobody outside OpenAI can use, which means nobody can reproduce them. The company says only that it is “working to responsibly release” it. The post gives no total for how many results exist.
Two concessions stand out. OpenAI is still looking for a home that meets the guidelines of an advisory group at the Institute for Advanced Study, so the GitHub repository it chose for now falls short. It also promises better citations and presentation in future releases, which admits this round is rough.
Lean is a programming language that lets a computer check a proof line by line. The repository carries Lean versions of many proofs, not all, with more promised. Any proof without one has been checked by no machine.
Until the model ships, treat the batch as OpenAI’s claim to scrutinize, not a capability to test.
OpenAI, “Sharing AI progress in mathematics,” published 6 October 2026.
