OpenAI Publishes 722 Math Manuscripts From Unreleased Frontier Model
The company published hundreds of research preprints and computer-checked Lean proof artifacts, igniting debate across the mathematical community.
Key takeaways
- OpenAI released 722 mathematical manuscripts grouped into 372 families in a public GitHub repository under an Apache-2.0 license.
- According to OpenAI, the proofs were generated by an unnamed, unreleased internal frontier model using an average of three hours of ChatGPT Pro thinking compute per result across approximately 4,000 attempted problems.
- Only a subset of the work is machine-checked, with the repository listing 162 papers with formalized main results in Lean.
- The release follows guidance and criticisms from the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study.
On October 6, 2026, OpenAI published a vast public repository of mathematical manuscripts generated by an unreleased internal frontier model. Hosted on GitHub under the openai/math repository, the collection includes 722 preprints organized into 372 distinct result families across 17 mathematical disciplines, accompanied by supporting proof artifacts and reasoning traces.
The release follows OpenAI's earlier announcement of an AI-generated solution to the Navier–Stokes existence and smoothness problem. According to the openai/math repository README, OpenAI expanded model evaluations to open research problems after its existing mathematical benchmarks reached saturation.
Inside the 722 Math Manuscripts and Research Scope
The repository catalogues results spanning number theory, algebraic geometry, theoretical computer science, combinatorics, differential equations, and mathematical physics. As reported by Kingy AI, the largest disciplinary categories are theoretical computer science with 40 families, combinatorics with 37, algebraic and complex geometry with 36, and number theory with 31.
According to the openai/math repository catalogue and analysis published by Kingy AI, prominent claims in the collection include:
- Number Theory: A claimed zero-free region for Dirichlet L-functions, including the Riemann zeta function, in the half-plane Re(s) > 7/8, alongside a companion argument for Re(s) > 11/12.
- Constants: A proof asserting that the irrationality exponent of π is exactly 2, and a separate claim that Catalan's constant is irrational.
- Algebraic Geometry: A proof of the rational Hodge conjecture for complex abelian varieties with complex multiplication across all dimensions and codimensions.
- Group Theory and Algebra: A resolution of the free group factor isomorphism problem and a proposed counterexample to Kaplansky's zero-divisor conjecture.
- Theoretical Computer Science: Claims addressing ordinary NP-hardness at the basic semidefinite threshold and reductions for Hilbert's tenth problem over the rational numbers.
OpenAI noted that two specific investigations departed from its standard automated pipeline: the Re(s) > 11/12 Riemann zeta zero-free region write-up, which was human-edited for readability, and the Hodge conjecture proof for CM abelian varieties.
Compute Usage and Generation Methods
OpenAI reports that the vast majority of manuscripts were produced using a uniform procedure. Across the evaluation, the internal model was posed roughly 4,000 research problems. An OpenAI spokesperson told Scientific American that the model produced almost every result in response to a single prompt given to a single AI agent, though some results may have required multiple attempts.
OpenAI reports that each published result required, on average, the equivalent compute of roughly three hours of ChatGPT Pro thinking. OpenAI disclosed 10 abridged reasoning summaries to illustrate the model's intermediate steps. However, as noted by Kingy AI, OpenAI did not disclose aggregate dollar costs, token volumes, or accelerator hardware hours, and it did not publish the specific prompts used for each problem.
Verification Status in Lean and Potential Errors
The repository contains artifacts written in Lean, a formal programming language that allows computer verification of mathematical proofs. However, formalization remains incomplete across the catalogue.
As detailed by OrcaRouter, the formalization manifest inside the repository lists 162 papers with a machine-checked main result, representing roughly one-fifth of the 722 manuscripts. While the table of contents links Lean documentation for 235 families, the metadata marks verification reviews as unchecked agent output with partial progress.
OpenAI acknowledged these limitations directly in its README, warning that some unformalized results could contain errors. The company stated it will record revisions and corrections as new versions while preserving previous release snapshots, and plans to add further Lean formalizations over time.
Community Response and Advisory Guidelines
The release follows public recommendations issued on September 29, 2026, by the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. After collecting more than 600 responses from mathematicians, as reported by Unite.AI, the group urged AI labs not to test open problems on proprietary models hidden from the research community, as reported by Scientific American.
The advisory group called for releasing results to independent repositories, publishing full prompt records and model names, formalizing proofs in Lean, and funding human-led efforts to parse AI outputs. While OpenAI published compute estimates, problem attempt counts, and reasoning summaries, it hosted the files on its own GitHub organization and has not released model weights or an API endpoint.
To help the research community evaluate the output, OpenAI announced it will fund a series of workshops, conferences, and special programs. The company stated it is working toward a responsible public release of the underlying frontier model, which remains unnamed and inaccessible outside internal evaluations.
Frequently asked questions
What did OpenAI release in the openai/math repository?
OpenAI published 722 mathematical manuscripts organized into 372 families under an Apache-2.0 license, including Lean code artifacts and 10 reasoning summaries from an internal frontier model.
Are all 722 mathematical manuscripts verified as correct?
No. The repository's formalization manifest lists 162 papers with machine-checked main results in Lean. OpenAI noted that unformalized papers could contain errors and has committed to publishing corrections as updated versions.
Can developers or researchers access the AI model that generated these proofs?
No. The model remains an unreleased, unnamed internal system. OpenAI stated it is working on a responsible release but has not provided a launch date, API endpoint, or pricing.
How much computing power was used per result?
OpenAI reports that each result consumed, on average, the equivalent compute of approximately three hours of ChatGPT Pro thinking, drawn from an evaluation batch of roughly 4,000 attempted problems.
Sources
- Sharing AI progress in mathematicsOpenAI · Official
- openai/mathGitHub · Oct 6, 2026 · Official
- OpenAI Releases 722 Math Manuscripts From an Unreleased AI ModelUnite.AI · Oct 6, 2026
- OpenAI Says Internal Model Produced 722 Math Manuscriptsfourweekmba.com · Oct 6, 2026
- OpenAI’s 722 Math Manuscripts: The Results, Proofs, Compute and CostsKingy AI · Oct 6, 2026
- OpenAI's 722 Math Manuscripts: The Model Behind Them Has No Name, No Price and No APIOrcaRouter · Oct 7, 2026
- OpenAI unleashes hundreds more math results upon a field already in shockScientific American · Oct 6, 2026
How this story was made: the newsroom picked it up from Google News and Reddit, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (25 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 7, 2026 at 01:42 UTC
