OpenAI said an unreleased internal AI model solved hundreds of open math problems using a single prompt. According to NS3.AI, the company published 722 manuscripts on GitHub on Tuesday, though that figure refers to manuscripts rather than solved problems.

An OpenAI spokesperson said almost everything came from one prompt given to one AI agent, although some results may have required multiple attempts. MIT mathematician Andrew Sutherland said the single-prompt, single-agent claims remain unverified until researchers can access the model and replicate the results.

Only 162 manuscripts have computer-checked main results, according to the repository's formalization catalog. OpenAI also warned that some unformalized results could have issues.