OpenAI publishes 722 math manuscripts

The headline: a 722-paper pull request for mathematics

OpenAI’s new mathematics release is not a normal research announcement. It is a repository-sized event: 722 manuscripts, grouped into 372 result families, produced by an unreleased internal OpenAI model and published for public inspection [2]. In its official post, OpenAI framed the batch as “a broad range of new mathematical results” from an internal frontier model, with Lean formalizations, compute estimates, reasoning summaries and revision protocols attached to the release [1].

That makes the story both breathtaking and unresolved. The strongest version of the claim circulating today is that the collection solves 90 of the top 500 open problems in mathematics, with the Quasi-Riemann Hypothesis among the marquee results [6]. But the careful reading is more cautious: OpenAI has published a large corpus of manuscripts and proof artifacts; the mathematics community now has to determine which claims survive formal, expert and reproducible review.

The metaphor almost writes itself: mathematics has received a 722-paper pull request. It may contain major theorems, elegant shortcuts and new proof techniques. It may also contain mistakes, over-broad statements and artifacts whose formal statement is narrower than the prose. The release is a milestone either way, because the bottleneck has shifted from generating candidate research to auditing it.

What OpenAI actually released

The public repository says the catalogue contains 722 manuscripts organized into 372 families, where a family can include a principal result, companion arguments, consequences or alternative proofs [2]. The material includes preprints, source files, manuscript-specific citation and build instructions, a manuscript map, a Lean library and a formalization catalogue [2].

OpenAI says the results came from evaluating models on open research problems after its existing mathematics evaluations saturated [2]. The company reports that approximately 4,000 problems were posed during the evaluation and that the average result used the equivalent of roughly three hours of ChatGPT Pro thinking compute [1][2]. OpenAI also says it is releasing ten abridged summaries of the model’s reasoning and statistics about attempted problems [1].

The model itself is not public. OpenAI says it is “working to responsibly release” the model that produced the results, but the current release is a research corpus, not a product launch, API endpoint or downloadable system [1]. That distinction matters: readers can inspect the manuscripts and some proof artifacts, but they cannot yet reproduce the discovery process from the model side.

Why the Quasi-Riemann claim matters

The most visible result family is the one commentators are calling the Quasi-Riemann Hypothesis. AINews described it as one of the central claims in the release and placed it near the top of the current reaction cycle, alongside discussion that many of the results correspond to major open problems [6].

A separate independent rerun focused on OpenAI’s Lean proof of a zeta-function zero-free region: the theorem checked there states that the Riemann zeta function is nonzero when the real part of the input is greater than 7/8 [7]. That is not the full Riemann Hypothesis, which concerns zeros on the critical line with real part 1/2, but it would still be a serious advance if the formal statement, its dependencies and its connection to the written manuscript all hold up [7].

The independent checker reported that Lean’s default kernel accepted the proof and that a second kernel, nanoda, accepted it too [7]. The same checker was explicit about limits: it reviewed the computer proof rather than the accompanying paper, and it was one independent run on one machine [7]. In other words, this is encouraging early evidence, not the end of the verification story.

The verification gap

The key number after 722 is 162. CellCog’s repository read found Lean formalizations for the main result of 162 papers, while Stanford Tech Review separately counted 162 unique manuscript folders in the formalization catalogue out of 722 manuscript folders in the map [4][5]. Stanford Tech Review put that at 22.4% of the manuscripts, leaving 560 without a formalized main result in that count [5].

OpenAI’s README itself warns that the collection includes results at different stages of verification, that not all manuscripts have accompanying Lean formalizations, and that some unformalized results could have issues [2]. That warning should be printed in bold on every discussion of the release. Lean can be a powerful check on a formal proof of a formal statement, but it does not automatically verify that a manuscript’s abstract, title and broader mathematical interpretation are all correct.

There is also a timing issue. Stanford Tech Review reported that none of the 143 manuscripts dated October 3 through October 6 appeared in the formalization catalogue it examined, while most of the Lean coverage was concentrated in late-September files [5]. That does not prove the late manuscripts are wrong; it may simply show that formalization lagged behind manuscript generation. But it reinforces the central point: the release is a queue for verification, not a final ledger of accepted theorems.

Community norms and the advisory question

OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study and drew on public recommendations while preparing the release [1]. The company also says it is exploring community-hosted alternatives for the materials and plans future improvements in citations, exposition and presentation [1].

That did not settle the process question. TokenPost reported that the release did not include every disclosure recommended by the advisory group, including placement in an academic repository not controlled by an AI lab and per-result disclosure of model name, prompts, reasoning summary, time spent and compute cost [8]. OpenAI’s repository is public, but it remains in OpenAI’s GitHub account; the model is unnamed and unreleased [1][8].

This is not a side issue. For mathematicians, the social machinery of publication matters because it tells readers what has been checked, by whom, and under what assumptions. If AI systems can generate hundreds of plausible manuscripts overnight, then provenance, versioning, independent hosting and reproducible proof checking become part of the theorem.

What to watch next

The next phase will be slower than the announcement. Mathematicians will triage the catalogue, rerun Lean builds, compare formal statements with manuscript claims, inspect dependencies and decide which results deserve seminar-level attention. OpenAI says corrections and revisions will be recorded as new versions and that previously released versions will remain accessible [2]. That version history may become as important as the original drop.

Three signals matter most now. First, independent reproduction: the zeta proof rerun is a useful start, but the community needs many more fresh clones, clean builds and independent kernel checks [7]. Second, expert review: formalization can certify a theorem statement, but mathematicians still need to assess novelty, significance and whether the written result says exactly what the formal artifact proves. Third, scope discipline: the phrase “90 of the top 500” is powerful, but rankings of open problems are not a canonical mathematical institution, so each headline claim should be traced to a specific manuscript and proof status [6].

The safest conclusion is also the most dramatic. OpenAI has not merely posted a benchmark score; it has posted a mountain of candidate mathematics. If even a fraction of the high-profile claims survives, machine-assisted theorem discovery has crossed a visible threshold. If many fail, the release will still define the new verification problem created by AI research at scale. Either way, mathematics now has a 722-manuscript audit on its hands.

Sources from the last 72 hours

1. [1]Sharing AI progress in mathematics | OpenAIOct 6, 2026, 2:00 PM
2. [2]math/README.md at main · openai/math · GitHubOct 6, 2026, 11:58 PM
3. [3]OpenAI Math Repository: 722 Manuscripts, 372 FamiliesOct 7, 2026, 8:29 AM
4. [4]OpenAI's 722 AI Math Papers: What's Proved, What's Checked | CellCogOct 6, 2026, 2:00 AM
5. [5]OpenAI's 722 AI Math Proofs: Only 162 Checked in LeanOct 7, 2026, 2:00 AM
6. [6][AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematicsOct 7, 2026, 2:00 AM
7. [7]davegoldblatt/openai-zeta-proof-check — ShadowgraphOct 7, 2026, 2:00 AM
8. [8]OpenAI Releases 722 AI-Generated Math Manuscripts After Criticism | TokenPostOct 7, 2026, 4:53 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.