Can AI do new math? What OpenAI's math release shows

What happened

On October 6, 2026, OpenAI, the company behind ChatGPT, put more than 700 math manuscripts online. A manuscript is a research paper that experts have not formally reviewed yet. They were written by an AI model that OpenAI uses internally and the public cannot use.

OpenAI says the model was given about 4,000 problems, including open problems: questions that mathematicians had not been able to solve. For each result, the model spent about three hours "thinking", which means working through the problem step by step before it writes an answer.

How the work is checked

In math, a result only counts once its proof has been checked. A proof is a step-by-step argument that shows a statement is always true. One wrong step can break the whole proof.

• Computer checks: for about 42% of the main results (as of October 7), OpenAI also shared a version written for Lean. Lean is a program that checks every step of a proof and rejects anything that does not follow. People still need to confirm that what Lean checked is the same claim the paper makes.
• Human checks: the other results have no computer check yet. Mathematicians are reading them, and several well-known mathematicians have said a result should not count until people have checked and understood it.

The first corrections

One day later, on October 7, OpenAI withdrew 3 manuscripts. A sign error (a plus where a minus belonged, or the other way round) in one paper broke an argument that two other papers relied on. 14 more manuscripts were corrected.

OpenAI itself describes the files as "raw research data" rather than finished papers.

Why it matters

The release shows that AI can now work on very hard problems for hours. It also shows the other side: even the strongest systems make mistakes that look small and change everything, and checking them takes experts a lot of time. How big a step this is for math is still being debated.

What this means for you when you chat with ChatPrime

When you ask ChatPrime to solve or explain a math problem step by step, the steps can look convincing and still contain an error.

• Check each step, not only the final answer.
• Put the answer back into the problem to see if it works, or solve it a second way.
• For homework, use the explanation to learn the method, then check the result yourself.

Source

OpenAI, math manuscripts, Lean files and change log, October 2026: github.com/openai/math

Corrections of October 7, 2026: history.md