The announcement: a batch of new math, not a single trophy result

On October 6, 2026, OpenAI published a page titled "Sharing AI progress in mathematics" announcing what it calls a broad range of new mathematical results produced by an internal frontier model. The announcement itself is a verified observed fact, captured on OpenAI's own site and reviewed by our editor. What the results actually prove, and whether the proofs are correct, is not independently verified as of publication. This article reports the announcement and its disclosure machinery, and keeps those two categories separate throughout.

The headline numbers, per OpenAI: formalizations of many of the proofs in Lean, a programming language that lets a computer check mathematical proofs; 10 summaries of the model's reasoning; estimates of compute spent expressed as ChatGPT Pro usage; and statistics about how many problems the system attempted. OpenAI says the average result used the equivalent of roughly three hours of ChatGPT Pro thinking.

What three hours of ChatGPT Pro thinking means

The exact compute sentence is material to this story, because it is the clearest public statement OpenAI has made about the cost of machine-produced research. From the October 6 release:

"The average result used the equivalent compute of roughly three hours of ChatGPT Pro thinking."

OpenAI, as the publisher of the release, October 6, 2026, page "Sharing AI progress in mathematics."

That unit deserves unpacking for non-specialists. A ChatGPT Pro subscription is OpenAI's paid consumer tier, and "thinking" refers to the model's extended reasoning time. OpenAI is not publishing a dollar figure or a kilowatt-hour figure here; it is translating its internal compute into a quantity a reader can price informally against their own subscription. That is an estimate supplied by the company that made the results, about its own workload, not an audited accounting. It is still far more cost detail than the previous release offered, and it makes one thing concrete: if the figure holds, machine-assisted research at this level is measured in hours of consumer-tier compute per result, not industrial-scale campaigns, at least as OpenAI describes it.

A repository with citation protocols, not a press release

The second headline choice is the container. Instead of a press release or a polished paper, OpenAI says it is publishing the results in a GitHub repository, with protocols for paper revisions and citations. Per the announcement, OpenAI is continuing to explore other community-hosted alternatives that meet the committee's guidelines, and commits to improving paper quality through citations, mathematical exposition, and presentation in future releases.

Why does a repository matter? Because a press release is a claim and a repository is an address. A GitHub repository is a versioned, public workspace where anyone can read files, see when they changed, and raise issues. With explicit protocols for revisions and citations, the mathematical community can reference specific results, flag errors, and track how the papers evolve, rather than reacting to a static announcement that cannot be amended in public view. This article could not independently open the repository itself (the host was outside this newsroom's retrieval allowlist at review time), so readers should treat the repository's contents as described by OpenAI and linked from its page, pending direct inspection.

The same caution applies to the committee. OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence (AGMAI) at the Institute for Advanced Study, and that the group's advice and public recommendations informed the release. The OpenAI page links to the group's site and to its public recommendations. We could not retrieve those pages directly, so we attribute every statement about AGMAI to OpenAI's description of it. The committee's independence is OpenAI's characterization; we have not confirmed the group's charter, membership, or funding from its own records.

What a Lean formalization actually buys you

The most technically interesting item is the Lean formalization. Lean is a programming language built so that mathematical proofs can be checked by a computer: a mathematician translates a proof into Lean's formal language, and a verification program accepts it only if every logical step follows from previously accepted foundations. There is no reviewer who nods along out of trust; the checker either compiles the proof or reports an error.

For a general reader, this changes the burden of proof. A human-written paper requires you to trust authors and referees. A Lean-checked proof requires you to trust a small, heavily scrutinized verification core and the honesty of the translation from the paper's argument into the formal language. That translation step is a real source of possible error or mismatch: a formalization can prove something subtly different from what the prose paper claims. So Lean verification is strong evidence, not absolute certainty, and OpenAI itself frames the formalizations as an ongoing effort, saying it will update the repository with more as they are obtained. Neither the formalizations' correctness nor their coverage of all announced results is independently verified here.

How this differs from the September Navier-Stokes release

The October 6 release is the second installment of a pattern. On September 8, 2026, OpenAI published a claimed solution to the Navier-Stokes existence and smoothness problem, one of the Millennium Prize Problems, including an analytical writeup and a Lean formalization. That earlier release described a multiagent effort involving on the order of 10,000 concurrent agents, 2.7 million agent messages and roughly 130 billion output tokens for the Navier-Stokes work alone, and an explicit decision not to claim the Millennium Prize. Our coverage of that single-result release is held as a separate, prior story; it appears here only as context for how OpenAI's disclosure practice evolved from one monumental claim to today's batched, protocol-driven release.

In the roughly four weeks between the two announcements, three things changed. The scope went from one Millennium Prize problem to a broad range of results. The documentation went from a paper plus formalization to a repository with revision and citation protocols, reasoning summaries, and compute estimates. And the process acquired an outside reference point: OpenAI says AGMAI's advice and public recommendations shaped how these results were released. Those are observable changes in OpenAI's own disclosure behavior, whatever one concludes about the underlying mathematics.

What to watch next

OpenAI's October 6 page makes two forward-looking commitments worth tracking because they can be checked later. First, it says it will fund a series of workshops, conferences, and special programs around the understanding of major results produced by AI, with more details to come. Funding the community that evaluates your claims is a meaningful gesture; the test will be whether the programs materialize and whether evaluation is independent of the funder.

Second, OpenAI says it is working to responsibly release the model that produced these results, and commits to continuing to evaluate internal frontier models on mathematics and other sciences and to act on community feedback in future disclosure standards. The model itself, its capabilities, and its release conditions are all unverified future events. Treating them as commitments to monitor, rather than facts, is the correct posture.

The optimistic reading, stated here as opinion: if OpenAI keeps this cadence, machine-contributed mathematics stops being a sequence of stunning headlines and becomes an auditable public record, checked by machines where possible and by the math community where not, with costs disclosed well enough to compare approaches. The skeptical reading: a vendor grades its own homework, the committee's independence is asserted rather than evidenced from primary records, and Lean formalizations cover only part of the claimed work. Both readings depend on facts this newsroom cannot yet independently confirm. The disclosure architecture is real and observable. The mathematics remains OpenAI's claim.