OpenAI Releases 722 AI-Generated Math Manuscripts—and Warns Some May Be Wrong

OpenAI published a 722-manuscript research collection on October 6

OpenAI’s official News feed records “Sharing AI progress in mathematics” at 12:00 UTC on October 6, 2026. Its description says an internal frontier model produced new results on open mathematics problems and that OpenAI released Lean proof formalizations and research details on GitHub. The public openai/math repository was created later that day, with initial commit adc7f1241b42e322a6451854ab7e4b4c146bf78a.

The release README counts 722 manuscripts grouped into 372 result families. Those units are different: a family may include a principal result, companion arguments, consequences or alternative proofs. The map ranges from number theory and algebraic geometry to analysis, theoretical computer science, mathematical physics and logic. This is not a model launch, API or generally available theorem-proving product; the model is explicitly described as internal and unreleased, and no price, API ID, access route or deployment date is supplied.

The production procedure was broad, fixed and compute-heavy

OpenAI says it expanded evaluation to open research problems after existing mathematical evaluations saturated. Approximately 4,000 problems were posed over the evaluation, and aggregating outputs by significance produced the published families and manuscripts. For the vast majority of results, OpenAI reports the same procedure and an average of three hours of ChatGPT Pro thinking compute per result.

That is a description from the producer, not an independently audited methods report. The release does not identify the internal model, publish a model card, expose sampling settings, provide a complete attempted-problem ledger, quantify unsuccessful attempts, or state a false-positive rate. Exceptions are also named for work on a Riemann-zeta zero-free region and the Hodge conjecture for CM abelian varieties, while one write-up was human-edited for readability. Readers should not infer one uniform provenance path for every file.

The headline claims are claims, not a new mathematical consensus

The manuscript map uses consequential language: entries say they prove or resolve problems including the quasi-Riemann hypothesis, Hilbert’s tenth problem over the rationals, the irrationality of Catalan’s constant, the Unique Games Conjecture and other long-standing questions. AccessAllGPT has not validated those statements, their novelty, their assumptions or their relation to prior literature. A repository entry and a polished PDF establish public availability, not correctness or priority.

OpenAI’s own README gives the essential caution: results sit at different stages of verification, not all have Lean formalizations, and some unformalized results could have issues. It says revisions and corrections will be released as new versions while older versions remain accessible. That version policy is useful, but future corrections do not retroactively validate the initial commit.

Formal artifacts improve auditability but do not flatten the evidence ladder

At the reviewed commit, AccessAllGPT counted 235 result-family documentation pages linked as Lean material, 162 source records in formalization.yaml and 405 JSON Comparator challenge configurations. These counts describe different layers and should not be equated. A family can have multiple manuscripts or proof targets; a source record can map to one or more declarations; a challenge configuration is an instruction for checking an artifact, not the result of an independent check.

OpenAI’s Lean README recommends compiling small portions because the combined library is large and warns that building everything can encounter Linux mmap limits. Comparator instructions require comparator, landrun and lean4export, then show a specific challenge command. Independent verification therefore means recording the exact commit, toolchain, dependencies, command, output, axioms and trusted computing base—not merely observing that a .lean or .json file exists.

The release exposes more than final PDFs

The repository includes an overview, a 372-family map, per-manuscript directories, PDFs, many TeX source trees, citation blocks, a Lean library and ten abridged reasoning summaries. The reasoning summaries cover selected results such as ordinary two-point correlations, the irrationality exponent of π, Mahler conjectures, Kaplansky direct finiteness, the Mézard–Parisi formula and relativistic Vlasov–Maxwell. Ten summaries are not a complete trace for 722 manuscripts.

The root repository carries an Apache 2.0 license and the README asks users to cite individual manuscripts with the BibTeX block in each directory. Before reuse, teams should still inspect the exact artifact and its notices rather than infer that every external reference, dependency or quoted prior result is relicensed. Citation also needs a commit or release version because OpenAI says corrections will create new versions.

A practical review starts with one bounded claim

Do not begin by asking whether “the 722 papers are correct.” Pick one result relevant to a real research program. Freeze the commit and manuscript hash; rewrite the theorem with every quantifier and hypothesis; map each dependency to prior literature; inspect whether the claimed novelty is actually new; and ask a specialist to look for scope changes, hidden regularity assumptions, circular references and gaps in cited lemmas.

If a Lean artifact is present, build only the named target first, inspect sorry counts and axioms, run the published Comparator challenge where applicable, and preserve logs. Then compare the formal statement with the natural-language theorem: a machine-checked declaration can be correct while formalizing a weaker or differently scoped claim. If no formal artifact exists, label the item “author-produced preprint, not independently verified” until substantive review changes that status.

What AI builders and research institutions should do next

Treat this release as evidence that frontier-model evaluation is moving beyond contest benchmarks toward large research corpora with auditable artifacts. It is not evidence that an available OpenAI product can autonomously solve open mathematics on demand. Builders should wait for model identity, access terms, evaluation protocol and failure-rate evidence before making product or procurement claims.

For the collection itself, adopt a claim-level triage process when a manuscript is relevant and independently reviewable; constrain public language to “OpenAI reports” until verification is complete; wait on high-consequence results lacking expert review or matching formalization; and reject any workflow that bulk-imports titles as established theorems. The most useful next signal will be transparent corrections, reproducible proof-checking reports and independent expert assessments—not the raw manuscript count.

Disclosures

AccessAllGPT received no OpenAI model access, ChatGPT Pro compute, repository preview, briefing, manuscript, proof artifact, review, payment or compensation for this article. OpenAI did not sponsor, review or endorse it. AccessAllGPT did not score or rank the mathematics and makes no finding that a listed theorem is true or false. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenAI, GitHub, Lean, mathlib or Comparator. Publication-wide relationships are listed on the disclosures page.