OpenAI Math Papers Release Splits Mathematicians Over Verification

Artificial intelligence has moved from solving single hard problems to producing mathematics in bulk. On Tuesday, October 6, 2026, OpenAI released 722 mathematical manuscripts generated by an internal frontier model it has not made public, putting them all in a GitHub repository. The OpenAI math papers cover hundreds of problems proved or disproved, and the reaction from mathematicians ranges from astonishment to deep scepticism about whether the results have been properly checked.

What happened

According to The Next Web, the 722 manuscripts fall into 372 families of results. The repository says OpenAI gave the model about 4,000 problems during its evaluation, and each result used on average about three hours of ChatGPT Pro thinking compute. The model is the same one behind OpenAI's Navier-Stokes proof last month.

Every earlier version of each paper will stay public, and each has its own citation block, the repository says. Many of the proofs come with formal versions in Lean, a programming language that lets a computer check a proof. Not all do, and OpenAI said some results without a formal version could have issues that it would fix quickly. OpenAI also published ten summaries of the model's reasoning. Two results did not follow the standard procedure: a zero-free region for the Riemann zeta function, whose write-up a human edited for readability, and a proof of the Hodge conjecture for CM abelian varieties.

New Scientist reported that OpenAI did not name the model, describing it only as an "internal frontier model." Nearly every paper came from a single prompt to a single AI agent, OpenAI told Scientific American, according to The Next Web. The company said it will fund workshops, conferences and programmes on AI-produced results and is working to release the model responsibly.

Why it matters

Some experts are amazed. Francis Johnson of University College London had worked for years on Wall's D(2) problem, one of the puzzles in the release. "I worked on this problem for 25 years. I produced two books on it. I, personally, gave up," he told New Scientist, adding that "the genie is out of the bottle now."

Others urge caution. Kevin Buzzard of Imperial College London said 30 papers related to his field of number theory, but only seven seemed impressive and only one was formally verified in Lean. "Acceptance of these results by the community will take time," he said. Tristan Buckmaster of New York University, who had worked on the Navier-Stokes problem, told The New York Times he doubted OpenAI had checked so many results. "I don't think they've done their sort of due diligence at all," he said. MIT's Andrew Sutherland told Scientific American to treat the single-agent claims as unverified until others can run the model.

Ben Allanach of the University of Cambridge told New Scientist that labs see mathematical discoveries as "an achievable intellectual trophy," and questioned whether results released at this scale, with little explanation, will really enter human knowledge. Unusually, the papers were published on GitHub rather than a preprint server; New Scientist noted that arXiv had just announced upload limits to stem "low-value" submissions.

The OpenAI math papers also clash with new guidance on how labs should share such work. The Advisory Group on Mathematics and Artificial Intelligence (AGMAI), hosted by the Institute for Advanced Study, issued guidelines on September 29 asking labs to stop testing hard problems on private models and to publish the prompts, time and compute behind each result. New Scientist said the release appears to go against several of those suggestions, including naming the model and disclosing how many comparable problems failed. AGMAI said on Tuesday night that its advice was not an endorsement and that "this release is the beginning, not the completion, of the process of human understanding."

OpenAI's Lindsay McCallum Rémy told New Scientist the company is "continuing to explore other community-hosted alternatives" that meet the committee's guidelines. The release adds to OpenAI's busy autumn, which has included enterprise projects such as the Ironclad study of GPT-6 Astra, while rivals like Google have also restricted access to their most capable systems, as with Gemini 4 Argon.

What's next

The checking now falls to human mathematicians, and that will take months. Buzzard said the unformalised results must wait until an expert reads them or a Lean formalisation is produced. Watch for formal verifications, retractions or corrections in the repository, and for whether OpenAI names the model or releases it to researchers. The OpenAI math papers could reshape how mathematical research is done, but only once the field has confirmed what is actually true in them.