Atlas AI Resources

← Back to AI News

AI research

OpenAI Posted 722 AI-Written Math Papers. Checking Them Will Take Months.

Published October 7, 2026

On Tuesday evening, October 6, OpenAI published 722 math papers written by an internal AI model it has not released to the public. The company says they cover 372 results, many of them answers or real progress on long-standing open problems. Many come with proofs a computer has checked, but mathematicians have only just started reading, and the release skipped some disclosures that an outside group of mathematicians had asked for.

What happened

The papers sit in a public GitHub repository (an online folder for code and documents) called openai/math. A result family groups a main result with companion papers, such as alternative proofs. OpenAI says it posed the model about 4,000 problems and kept the outputs that met a significance bar. On average, each result used the equivalent of about three hours of ChatGPT Pro “thinking” with the internal model. OpenAI says it already tested its models on open research problems and expanded that testing after they maxed out its usual math tests.

The claimed results include a four-dimensional version of the Kakeya conjecture (a famous question about how thin a shape can be while still containing a line segment pointing in every direction) and a step toward the Riemann hypothesis, the most famous unsolved problem about prime numbers. That step does not prove the hypothesis. It rules out the hidden “zeros” the hypothesis is about from a wide band, while the full hypothesis would pin all of them to a single line. OpenAI also notes that this Riemann work, and one other result, did not follow the standard procedure used for the rest, and that one write-up was edited by humans for readability.

Many papers come with proofs in Lean, a programming language in which a computer checks every step of a proof. In the repository’s own index, 235 of the 372 families link to Lean material. OpenAI also shared ten shortened summaries of the model’s reasoning, and it says some unchecked results “could have issues.”

Why it matters to a normal person

Math is one of the few places where an AI’s claims can be checked strictly, so it has become a public test of how capable these models are. If even a fraction of these results hold up, that is a real sign that AI can now do research-level reasoning, not just look things up.

It also raises questions of fairness and credit. An advisory group of mathematicians hosted at the Institute for Advanced Study asked AI labs on September 29 to release results in a repository the lab does not control, with the prompts, costs, and failed attempts, and to stop testing hard problems on models nobody else can use. Scientific American reports that OpenAI shared only average compute figures and no prompts, and an OpenAI spokesperson said the company is not bound by the group’s advice. The group said on Tuesday that its role “should not be interpreted” as an endorsement, and that the release is “the beginning, not the completion” of human understanding.

What it does not mean

It does not mean hundreds of problems are settled. Scientific American reports that sorting through the papers will take months, including deciding whether the proofs contain new ideas or mostly recombine known ones. The spokesperson also said many results are not yet understood by OpenAI’s own mathematicians.

A Lean check is strong evidence, but it confirms that a proof matches the formal statement, not that the formal statement says exactly what the paper claims. People still have to check that match.

And it does not mean anyone can do this today. The model is not public. OpenAI told Scientific American that almost every result came from a single prompt to a single AI agent, while adding that some may have taken several attempts. “We should ask for receipts,” MIT mathematician Andrew Sutherland told the magazine. Daniel Litt of the University of Toronto took the opposite view, saying the release will be “a good thing for mathematics.”

Sources