OpenAI says an unreleased internal AI model produced solutions to hundreds of math problems after receiving a single prompt, a claim highlighted by NS3.AI, which notes 722 manuscripts were published on GitHub on Tuesday. The 722 figure refers to manuscripts, not solved problems.

An OpenAI spokesperson indicated that nearly all results came from one prompt given to one AI agent, though some results may have required multiple attempts.

MIT mathematician Andrew Sutherland cautioned that the single-prompt, single-agent claim remains unverified until researchers can access the model and replicate the results.

The repository's formalization catalog shows that 162 manuscripts include computer-checked main results, and OpenAI warned that some unformalized results could have issues.