OpenAI releases 722 math manuscripts from an unreleased frontier model, arranging the work into 372 result families across multiple mathematical disciplines. The October 6 publication expands the evidence available for evaluating the model beyond OpenAI's earlier Navier–Stokes claim.

 

The release has four defining features:

  • 722 manuscripts grouped into 372 related families
  • Public PDFs, source files and citation instructions on GitHub
  • Lean formalizations for many, but not all, results
  • Ten abridged summaries of the model's reasoning

 

OpenAI Releases 722 Math Manuscripts From 4,000 Problems

OpenAI says it presented roughly 4,000 open research problems to the internal model during the evaluation. Results judged significant enough for publication were consolidated into families containing a principal result, companion arguments, consequences or alternative proofs.

 

The catalogue covers 372 families rather than 722 independent discoveries. That distinction matters because several manuscripts can address different parts of one mathematical result. The headline manuscript count measures the size of the collection, while the family count is the closer guide to the number of grouped research contributions.

 

The repository includes an overview, a manuscript map, preprint directories and an Apache 2.0 license. Individual folders contain PDFs, source material, bibliography information and build instructions, giving researchers a defined route to inspect, cite and revise each paper.

 

OpenAI reports that the average result used computing equivalent to about three hours of ChatGPT Pro thinking with the unreleased model. That is a standardized estimate, not a dollar cost or a guarantee that another system could reproduce the work with the same resources.

 

Lean Proofs Create a Stronger Verification Path

Many manuscripts are accompanied by formal proofs written in Lean, a language that lets a computer check whether every encoded step follows from declared definitions and assumptions. OpenAI has also published a catalogue connecting formalizations to their corresponding papers and instructions for additional comparator checks.

 

The repository explicitly warns that not every manuscript has been formalized and that some unformalized results may contain errors. OpenAI says it will add more Lean work and record corrections as new versions while preserving earlier public versions.

 

Formal verification narrows one class of risk but does not complete scholarly review. A checker can validate an encoded theorem while mathematicians still need to confirm that the statement captures the intended open problem, that prior work is credited properly and that the argument contributes useful understanding rather than only a correct certificate.

 

The ten published reasoning summaries offer another inspection layer. They cover topics including the irrationality exponent of pi, Mahler conjectures, spin glasses, free group factors and the relativistic Vlasov–Maxwell system, but they represent a small sample of the full collection.

 

The Mathematics Community Faces a Verification Bottleneck

The volume makes evaluation the immediate challenge. Reading 722 manuscripts, checking references, testing formalizations and determining novelty will require coordinated work across specialties. A public repository makes scrutiny possible, but it does not turn every uploaded paper into an independently accepted result.

 

OpenAI consulted the independent Advisory Group on Mathematics and Artificial Intelligence before publication. The group has urged AI laboratories to use established academic channels where possible, disclose models, prompts and computing costs, and avoid treating mathematical results primarily as model marketing.

 

For this release, OpenAI chose GitHub with revision and citation protocols while saying it is exploring community-hosted alternatives. The company also acknowledged that future releases need stronger citations, exposition and presentation, which are central to helping researchers understand how a result fits into an existing field.

 

The Verge reported that the collection includes solutions to hundreds of open questions, but the full impact will take time to establish. That cautious framing is essential: the repository is evidence that the model generated a large research corpus, not independent confirmation that every claimed result is new and correct.

 

The Frontier Model Remains Unreleased

OpenAI has not released the model that produced most of the manuscripts. It says it is working toward a responsible release, leaving outside teams able to inspect the outputs but unable to rerun the same model or determine how consistently it succeeds across new problems.

 

The company also identifies exceptions to its standard evaluation procedure, including work involving a zero-free region for the Riemann zeta function and the Hodge Conjecture for CM abelian varieties. One zeta-function manuscript received human editing for readability, an example of why provenance details matter at the paper level.

 

The next credible milestones will come from independent mathematical review, reproducible Lean builds, documented corrections and publication through venues that specialists trust. Those checks will show which families become durable additions to mathematics and which require substantial revision.

 

If a meaningful share survives scrutiny, the release will demonstrate a new research workflow: frontier models generating candidate mathematics at scale, formal tools checking portions of it and human communities deciding novelty, context and significance. The unresolved issue is whether verification capacity can expand fast enough to match the rate of machine-generated claims.

 

More on This Story