OpenAI posts 722 maths papers by an unreleased model, withdraws three

OpenAI published 722 mathematics papers written by an unreleased internal model on the 6th of October 2026, according to its math repository on GitHub. It withdrew three of them in an update that reached the repository on the 8th of October.

The collection now lists 719 papers, the front page says.

Why it matters

OpenAI wrote that its models now reach the top of its usual math tests. So it also tests them on open research problems.

One sign error in one paper also broke two papers that built on it, the repository's history page says.

Key facts

The model was given about 4,000 problems, the front page says. Each result used on average three hours of ChatGPT Pro thinking time.

Results include work on the Riemann zeta function and the Hodge Conjecture, according to the page.

The same update fixed 14 other papers, the history page says.

About 42 percent of the main results have been checked by Lean, a program that tests proofs by computer, the front page says. "Some of the unformalized results could have issues," it adds.

What's next

OpenAI said it will record fixes as new versions and keep the old ones online. The repository had more than 10,000 stars, GitHub's bookmarks, by the 8th of October.