Net Good IndexSubmit a correction

Benefit Ledger · provisional · Mathematics

OpenAI releases 722 mathematics manuscripts from an unreleased internal model

On 6 October 2026 OpenAI published 722 mathematics manuscripts grouped into 372 result families, produced mostly by an unreleased internal frontier model; some proofs have Lean formalizations, while others remain unverified.

Tier 4 HistoricMethodology 0.2
Submit a correction

Current score

+3.83

30 base · Historic (tier 4 of 5, 30 pts)
× 0.9000 attribution · Primary causal contribution
× 0.3500 evidence · Technical disclosure or preprint
× 0.6000 realization · Complete public manuscript
× 0.9000 durability · Effectively permanent
Event-level product before credit split: 5.10

A public corpus of 722 manuscripts across 372 result families is a Historic-scale mathematical research contribution if its central results withstand scrutiny (tier 4). OpenAI attributes the vast majority to the same unreleased internal research system already represented in the ledger (0.90); credit share is 0.75 to reflect human selection, grouping, editing exceptions, and the fact that this aggregate publication does not claim every result as exclusively model-produced. Evidence is a detailed public corpus and official account, not independent verification of the mathematics (0.35). Realization is a complete public manuscript collection with partial Lean formalization, not broad expert acceptance (0.60). Correct mathematical results are durable (0.90).

What happened

On 6 October 2026, OpenAI released a public collection of 722 mathematical manuscripts grouped into 372 families of related results. OpenAI says the vast majority were produced through a common procedure using an unreleased internal model; approximately 4,000 problems were posed, with about three hours of ChatGPT Pro thinking compute per result on average. The repository includes preprints, source files, citation and revision protocols, and Lean formalizations for many but not all proofs. OpenAI says some unformalized results could have issues and that additional formalizations will be added. The release includes claims about major open problems across mathematics, but this ledger entry treats the corpus as one provisional cluster and does not establish that every manuscript is correct or independently novel. Exceptions named by the repository include the Riemann zeta zero-free-region work and the Hodge conjecture result for CM abelian varieties; the 11/12 zero-free-region writeup was human-edited. Earlier individually recorded results from the same internal model, including the Navier–Stokes result, remain separate records and are not counted here as newly established theorems.

Model attribution

GPT-6

Unreleased internal frontier model credited by OpenAI with producing the vast majority of the mathematics manuscripts in the October 2026 collection.

OpenAI's official announcement and the repository identify an unreleased internal frontier model as the source of the broad collection. The repository names exceptions to the common generation procedure, and OpenAI researchers selected and grouped the manuscripts; individual earlier milestones remain separately represented.

Attribution 0.9000 · Credit share 75% · OpenAI

Claims

  • OpenAI published 722 mathematics manuscripts grouped into 372 result families in a public GitHub repository on 6 October 2026.

    outcome · supported

  • OpenAI says the vast majority of the collection was produced with an unreleased internal model after posing approximately 4,000 problems.

    attribution · supported

  • Every manuscript in the collection has been independently verified and accepted by the mathematical community.

    significance · disputed

  • All 722 manuscripts are accompanied by Lean formalizations.

    outcome · disputed

Sources

primary sources

other sources

Secondary domains: Education & Knowledge Access, Computer Science

Revision history

  • 7 Oct 2026 · 0.00 → 3.83

    Imported events under methodology 0.2.