OpenAI published a batch of mathematical results produced by an unreleased internal frontier model on its official blog on October 6, 2026, at a scale unusual for a single release: 722 manuscripts spanning 372 families of results, covering hundreds of long-standing unsolved problems. The work did not go to a journal. It was placed in a single GitHub repository, with rules for manuscript revision and citation, for the mathematics community to check.
What the release actually claims
In September 2026, OpenAI had already said its models had solved more than a hundred long-open problems across most branches of mathematics. This release turns that claim into readable text: alongside the proofs, it includes summaries of parts of the model's reasoning, estimates of compute used, and a count of how many problems the model attempted in total. By OpenAI's account, a typical result consumed compute roughly equal to ChatGPT Pro thinking continuously for three hours.
In other words, this was not a model having one lucky insight on one problem. It was a pipeline run at scale: large numbers of open problems fed one by one to the same unreleased model, with the output collected into manuscripts and released together. The volume itself is the signal — frontier models are moving from contest problems to systematically working through open research questions.
Why Lean formalization is attached
A substantial share of the proofs in this batch were formalized in Lean, a system that lets a computer check a proof line by line. Once formalized, correctness no longer rests on a reviewer's intuition; a machine can verify it. Trust is exactly where AI-generated work faces the hardest questions, and machine-checkable proofs are what set these manuscripts apart from ordinary preprints.
The boundary matters too: only part of the batch is formalized. The rest still needs human mathematicians to evaluate and absorb, and OpenAI itself notes that the full impact will take time to become clear. How many of the 722 manuscripts survive peer scrutiny is an open question for now.
The argument is about how results are released, not the answers
Discussion around this batch has focused less on the mathematics than on the method of release. In late September, AGMAI, an independent advisory group of mathematicians, issued its first recommendations: labs should publish mathematical results promptly through established academic channels, disclose the model name, prompts and compute used, and — in the group's pointed phrasing — should not treat publishing math results as a marketing vehicle for their models.
Measured against those points, this release answers some of them: the results are public in one place, compute figures are given, and revision and citation rules are included. The most important item is still missing — the model that produced the work has no name, because it has not been released. Readers can verify the proofs but cannot reproduce the process that generated them. For mathematics, a proof can be checked independently. For the AI industry, the thing worth watching is different: as models start producing verifiable research in bulk, whether labs' publishing norms can keep up with the output.
