OpenAI publishes 722 mathematical manuscripts on GitHub

The release groups machine-generated proofs across 372 research families while an independent advisory body urges public labs to disclose models and prompts.

OpenAI published 722 mathematical manuscripts in a public GitHub repository on October 6, grouping the machine-generated papers into 372 research families across multiple disciplines.¹² The company credited the proofs to an unreleased internal model that evaluated approximately 4,000 open problems.²¹ The public repository, issued under an Apache 2.0 open-source license, contains preprints, source files, and formal verification scripts spanning number theory, combinatorics, theoretical computer science, and mathematical physics.³²¹

The release represents an expansion of automated reasoning evaluations initiated after the company's performance on existing mathematical benchmarks saturated, according to repository documentation.²⁴ OpenAI reported that producing each surviving result required an average of three hours of ChatGPT Pro thinking compute.²¹ Nearly every paper resulted from a single prompt provided to an individual artificial intelligence agent operating under a standardized pipeline.⁵²

Automated proofs and uneven verification

Verification across the collection remains uneven.¹ Many manuscripts include formal proofs formatted for the Lean proof assistant, which allows a computer program to verify logical derivations step by step.¹⁵ OpenAI acknowledged in its repository notes that not all papers currently carry Lean formalizations, cautioning that unformalized manuscripts could contain errors that the company will attempt to repair as they surface.²¹ The repository maintainers pledged to update the catalogue with additional formalizations as they become available.²¹

Alongside the preprints, OpenAI published 10 abridged summaries outlining the reasoning traces behind selected results.²¹ The summaries describe solutions across varied mathematical subfields, including the irrationality exponent of pi, the symmetric and general Mahler conjectures, ordinary NP-hardness at the basic semidefinite threshold, and quasipolynomial bounds for arithmetic progressions.²⁶ Other summaries address Kaplansky's direct-finiteness conjecture in characteristic two, the Mézard-Parisi formula for diluted spin glasses, spontaneous magnetization in the quantum Heisenberg ferromagnet, isomorphism of free group factors, and the three-dimensional relativistic Vlasov-Maxwell system.²⁶

Two results deviated from the automated workflow.²¹ Work on the Hodge Conjecture for CM abelian varieties and an analysis establishing a zero-free region for the Riemann zeta function followed separate investigative tracks.²⁴ For the zeta function manuscript, OpenAI noted that the text analyzing the zero-free region was edited by human researchers to improve clarity and readability.²¹ The repository includes an explicit versioning policy designed to preserve the public release history, ensuring earlier iterations remain accessible whenever revisions appear.²¹

Academic pushback and release standards

The October release follows earlier mathematical announcements that prompted sharp debate within the academic community.¹⁵ On August 1, OpenAI unveiled an internal model named Astra with 10 results on long-standing problems, including an improvement to a 1978 sphere-packing bound.³ On September 8, the company announced that its systems had identified finite-time blow-up solutions to the Navier-Stokes equations, accompanied by a 166-page manuscript and a 57-page paper on the Euler equations.³ OpenAI previously disclosed that the Navier-Stokes effort deployed more than 10,000 artificial intelligence agents and cost several million dollars in compute.⁵ That announcement prompted 25 Fields Medalists to sign a joint declaration titled "A Severe Misalignment of AI in Mathematics," which criticized corporate releases that circumvent human peer review and omit individual attribution trails.³

The 722 manuscripts also arrived one week after the Advisory Group on Mathematics and Artificial Intelligence issued formal release guidelines.⁵¹ The nine-member panel, established under the Institute for Advanced Study in Princeton and serving without compensation, gathered more than 600 survey responses from professional mathematicians before publishing its recommendations on September 29.⁵¹ The advisory group urged research laboratories to deposit artificial intelligence outputs in independent academic repositories rather than corporate accounts.⁵ The guidelines also called on companies to disclose model identities, prompts, reasoning traces, elapsed runtimes, and compute costs for every paper, while warning against treating mathematical proofs as marketing campaigns.⁵

OpenAI chose to release the collection on its own GitHub repository, although documentation notes that the company is exploring community-hosted alternatives.²⁵ The company did not reveal the name of the model that produced the 722 manuscripts, omitted the prompts used for individual results, and set no timeline or access terms for public availability of the underlying system.¹ The advisory group clarified that its consultation did not constitute an endorsement of the manuscripts or OpenAI's release strategy, stating that publication marks only the start of academic evaluation.⁵

Working mathematicians have expressed divergent views on how the field should absorb large volumes of automated literature. Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, said claims regarding the manuscripts should be treated as unverified until the model itself is made public and the experiments can be independently replicated.⁵ Northwestern University mathematician Bryna Kra, who helped draft the Leiden Declaration signed by more than 4,000 mathematicians, said participants in an August gathering urged corporate labs to publish detailed papers that researchers can evaluate rather than relying on brief online posts.⁵ Daniel Litt, a mathematician at the University of Toronto, offered a different perspective, remarking that commercial companies have no practical reason to keep mathematical answers secret.⁵

What this rests on

32 sentences trace to 6 sources.

All 6Documents 0Records 0Reporting 6People 0

1. 1 R runtimewire.com Third-party reporting · runtimewire.com · 6 Oct 2026 OpenAI publishes 722 math manuscripts generated by an unreleased model See the source
2. 2 G github.com Third-party reporting · github.com math/README.md at main · openai/math See the source
3. 3 SF Startup Fortune Third-party reporting · startupfortune.com · 7 Oct 2026 OpenAI drops 722 AI math proofs and mathematicians are not impressed - Startup Fortune See the source
4. 4 FP Free Press Journal Third-party reporting · freepressjournal.in · 7 Oct 2026 OpenAI Publishes 722 AI-Generated Math Manuscripts From An Unreleased Model: Here See the source
5. 5 T tokenpost.com Third-party reporting · tokenpost.com · 7 Oct 2026 OpenAI Releases 722 AI-Generated Math Manuscripts After Criticism | TokenPost See the source
6. 6 G github.com Third-party reporting · github.com GitHub - openai/math See the source

Also consulted: 1 page across 1 site, none cited.