On Wednesday, October 7, 2026, OpenAI released a huge batch of AI-generated mathematical results in one go. According to Engadget, the release consists of 722 manuscripts hosted in a GitHub repository that address 372 open mathematical problems. New Scientist described it as 722 papers proving and disproving a wide range of statements, a scale that the academics it consulted called “as impressive as it is baffling.”
The company did not identify the system that produced the results, describing it only as an “internal frontier model.” According to The Decoder, each result is meant to solve an open problem or make substantial progress toward a solution. Among the achievements OpenAI claims are a solution to the four-dimensional Kakeya conjecture, improvements to important computer algorithms and advances related to the Riemann hypothesis, one of the most famous problems in mathematics.
It is worth stressing from the outset that these are the company's claims and that verification is under way. None of these results has yet gone through the usual peer-review process, and several experts insist that the community will need time to determine which ones are correct, which are original and which have genuine mathematical relevance. A logically valid result is not necessarily an important one.
The working method is also striking. OpenAI says nearly every manuscript came from a single prompt to a single AI agent, although some needed several attempts. On average, each result consumed about three hours of compute equivalent to ChatGPT Pro's reasoning mode. The contrast with the Navier-Stokes work OpenAI announced in September is notable: that effort required, according to The Decoder, a swarm of 10,000 agents and millions of dollars in compute, and it has been under formal review for weeks.
Part of the material includes formalizations in Lean, a programming language designed to check mathematical proofs automatically, and OpenAI has promised to add more. The reason is practical: the volume of results could overwhelm mathematicians' capacity for manual review. A Lean formalization, however, guarantees the logical correctness of a proof but says nothing about its originality or about whether the statement proved is the one the community actually cares about.
Publishing on GitHub rather than in scientific journals is itself a statement of intent. OpenAI explained that the repository has protocols for revising papers and handling citations, and that it is still exploring “other community-hosted alternatives” that meet the advisory committee's guidelines. Spokesperson Lindsay McCallum Rémy told New Scientist that the company wants “to give the mathematical community time and space to assess this work.”
That committee is the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), an independent group set up after Terence Tao of the University of California, Los Angeles, criticized last month the way discoveries were being “dumped” on the community. According to The Decoder, the group is linked to the Institute for Advanced Study and includes Fields medalist Timothy Gowers. Its recommendations call for disclosing the model's name, the prompts used, the compute cost and how many comparable problems were attempted without success.
OpenAI maintains that it followed those guidelines, but several outlets point to the opposite on specific points. Engadget notes that the company did not publish its prompts or per-problem compute times, offering only averages. New Scientist adds that the model has no public name either. The company did publish summaries of the reasoning, statistics on how many problems the model attempted and cost estimates, and it announced that it will fund workshops and conferences to study AI-produced results.
Mathematicians' reactions vary. Francis Johnson of University College London worked for years on Wall's D(2) problem, one of those included in the batch. “I worked on this problem for 25 years. I produced two books on it. I, personally, gave up,” he told New Scientist. Johnson said he was surprised by the speed, though not by the fact itself, and summed it up: “The genie is out of the bottle now.”
Kevin Buzzard of Imperial College London urged caution. As he explained to New Scientist, the release included 30 papers relevant to his field, number theory, but only seven struck him as impressive and only one had been formally verified in Lean. “Acceptance of these results by the community will take time,” he warned. MIT mathematician Andrew Sutherland was blunter with Scientific American: until the model is released and others can replicate the results, any such claims should be treated as unverified.
The debate goes beyond technical correctness. In a recent open letter titled “A Severe Misalignment of AI in Mathematics,” 25 Fields medalists warned that problem-solving is only a means toward conceptual understanding, and that mass-producing true statements could impoverish the field rather than enrich it. Ben Allanach of the University of Cambridge criticized the scale of the release and the limited effort to explain the results, which he described as “confusing and disrupting.”
The publishing context does not help either. New Scientist noted that, this same week, the preprint server arXiv announced upload limits to stem a tide of “low-value” submissions. If AI labs can generate hundreds of manuscripts within days, traditional review mechanisms, which depend on specialists' volunteer time, face a problem of scale. Gowers had already warned that within one or two decades the mathematical literature could keep growing without anyone truly understanding it.
For the industry, the announcement is another sign that mathematics has become a showcase for capabilities. Allanach described it as an “achievable intellectual trophy” for the companies. OpenAI says it is working on a responsible release of the model to “directly empower scientists.” If that happens, independent replication will become possible and the conversation will shift from promises to checks.
In short, the release of 722 manuscripts on 372 problems is a milestone because of its sheer volume, but its real value will depend on a verification process that is only beginning. The Lean formalizations help, the views of experts such as Johnson show that some results may be genuine, and the doubts voiced by Buzzard and Sutherland are a reminder that science is not settled by press release. The prudent framing, for now, is announced results with verification under way.
