Evidence at a glance
The Release Is a Set of Results, Not a Model
On October 6, 2026, OpenAI released a collection of mathematical results produced by an internal frontier model, placing draft papers, supporting materials, and some Lean formalizations in a GitHub repository. The release concerns progress on open research problems, not a model launch: the model itself remains unavailable, while outsiders receive organized results and part of their supporting evidence.
That distinction shapes how the collection should be read. OpenAI says it turned to open problems after existing mathematics evaluations began to saturate, then organized model outputs into result families and papers and selected work meeting an importance threshold. The public materials do not define that threshold or explain how each result was selected. This is therefore not a scorecard of model capability, but a collection of research claims that must be understood, checked, and assessed one by one.
The Numbers Show Scale, Not a Success Rate
The repository’s scale figures need to be separated before they can be interpreted: its listed contents include 722 papers grouped into 372 result families, and OpenAI says the model was asked roughly 4,000 problems during evaluation. Papers and result families are different units. A family may contain a main result, related arguments, corollaries, or alternative proofs, while the 4,000 figure counts attempted problems, not solved ones. Combining these numbers into a single “success rate” would misrepresent what the release demonstrates.
OpenAI also estimates that an average result used compute equivalent to roughly three hours of ChatGPT Pro thinking, and published reasoning summaries for 10 results. The first is a compute estimate, not a precise runtime or per-result cost. The summaries are not complete reasoning records. These disclosures offer some reference points for effort and process, but they cannot independently reconstruct the experiments or establish the quality and novelty of the results.
Lean Can Check Proofs, Not Make the Community’s Judgments
Lean formalization adds a layer of computer-checkable evidence to this research. Once a proof is expressed in a form Lean can interpret, the proof assistant can check whether the formal argument holds. That gives reviewers a more explicit verification route than a natural-language proof alone. OpenAI says many papers include formalized proofs and that more will be added, but it provides no total count, and not every paper has been formalized.
Even a proof that passes Lean leaves other mathematical questions open: whether the formalization faithfully captures the paper’s claim, whether the result is novel and important, and whether the definitions and assumptions are appropriate. OpenAI warns that papers without formalization may contain problems, while the public materials do not provide independent verification findings for every paper. Lean can check a proof within a formal system. It does not automatically perform peer review or decide which results belong in the body of mathematical knowledge.
More Transparency, with the Lab Still Behind the Curtain
OpenAI has released more than conclusions: the repository includes paper revision and citation protocols, selected reasoning summaries, compute estimates, and counts of attempted problems. These materials give external researchers shared objects for discussion and respond to concerns about the time needed to understand, check, and absorb AI-generated results. OpenAI says it sought advice from an advisory group on mathematics and AI before publication, but the group has explicitly said that advising does not mean endorsing the results or the release approach.
Even so, outsiders cannot fully reproduce the research process from the materials provided. The model’s specific name, prompts, and complete experimental workflow are not public, and the criteria for selecting results are unspecified. The reasoning summaries are summaries, not step-by-step records. The advisory group had recommended formalizing results where possible and disclosing the model, prompts, time, and compute cost. Comparing those recommendations with this release shows both progress and the information gaps that remain.
Treat the Collection as a Starting Point, Not a Verdict
For technical leaders and research teams, the practical lesson is that the quality of an AI research release depends on more than whether a model produces a correct answer. Outsiders also need to locate the claim, inspect the proof, track revisions, and understand where the result came from. This repository provides some of that infrastructure, particularly Lean files and citation and revision protocols. But with the model unavailable and the workflow incomplete, it cannot be treated as a reproducible research handoff.
A more reliable approach is to treat the papers as research leads: examine available formalizations first, then rely on domain researchers to assess novelty, importance, and scope. OpenAI’s work on a zero-free region for the Riemann zeta function also shows that the collection is not uniform. One paper concerning the region Re(s)>11/12 was edited by a human for readability, and the repository describes other work as following exceptional processes. The actionable conclusion is not that AI has independently proved important theorems. It is to ask, result by result, whether the proof is checkable, how much of the research process is disclosed, and what still requires independent review.