OpenAI posted 721 AI-written maths papers; by its count, about 42% have computer-checked proofs
An unnamed OpenAI model was set about 4,000 open problems and wrote up the results it claims. Most still await review by human experts.
On 6 October 2026 OpenAI published openai/math, a public collection of maths research papers written by an unnamed in-house AI model. OpenAI says the model was set about 4,000 open research problems and used about 3 hours of ChatGPT Pro thinking compute per result on average. It calls the papers raw research data and leaves questions of credit to mathematicians.
The first version had 722 papers. On 7 October a sign error forced OpenAI to withdraw three, related to the Hodge conjecture. It then reported 300 of the remaining 719 headline results (about 42%) checked in Lean, a language in which a computer verifies every step. Two papers added on 26 October brought the total to 721, in 373 families of related results.
The claims span 17 fields and include the quasi-Riemann hypothesis and a counterexample to Hadwiger’s conjecture. Epoch AI credits the release with solving 12 of the 68 open problems on its FrontierMath Erdős benchmark.
Reactions were mixed. Mathematician Matt Baker said the claims could be a watershed, but that the release should have been far better coordinated. Scientific American reported that some physics write-ups were judged nearly impossible to verify. Terence Tao, in paraphrased reporting, said it overtook junior researchers’ projects without warning. A community hub launched via his blog on 24 October to rewrite the papers for readability and proper citations.
What it means
One AI model has produced hundreds of claimed research results at once, far more than mathematicians have so far been able to check. About 42% come with a computer-checked proof; the rest, and whether each result matters, still need human experts. The work of checking, rewriting and crediting falls on mathematicians, including junior researchers whose projects it overtook.
Caveats
- By OpenAI’s own count, more than half of the headline results have no Lean proof, and its README warns that some of those could have problems. (The contents file links a Lean proof for 245 of 373 families, a different measure.)
- Lean certifies the formal statement, not that it matches the paper’s English claim.
- OpenAI has not released full prompts or failure rates. Of about 4,000 problems posed, only the successes are listed.
- Headline claims such as the Hadwiger counterexample and the Yang–Mills-related work have no independent confirmation in the sources we read.
- Counts differ between reports (722, 719 or 721 papers; 372 or 373 families; Quanta says 377 results). The repository has issues disabled and has locked pull-request discussions since 12 October.