OpenAI pushed a new batch of mathematical manuscripts to GitHub on October 6, 2026, crediting the work to an unreleased internal model. Two days later, the story has split into two numbers nobody can reconcile: the live repository lists 719 manuscripts sorted into 372 result families, while multiple outlets reported the original drop at 722 manuscripts. OpenAI has not explained the gap, and as of October 8 it remains unresolved. For a release built on the premise of mathematical rigor, a disputed head count is an awkward way to start.
What OpenAI Actually Posted on GitHub
The repository holds a collection of OpenAI math papers covering pure mathematics and theoretical computer science, generated by a model OpenAI has not named and has not released to the public. According to the announcement, the model worked through roughly 4,000 problems during its evaluation period, and the manuscripts published this week represent a curated subset of that output. Supporting proof artifacts accompany many of the papers, and a number of them include formal Lean proofs meant to let outside reviewers check the logic step by step rather than take the model’s word for it.
OpenAI itself has been careful to frame the release as a work in progress. The company said the published results sit at different stages of verification, which is a carefully hedged way to describe a repository branded as a mathematics breakthrough. That hedge matters, because it is the clearest signal yet that OpenAI expects some fraction of these claims to be revised, corrected, or dropped once mathematicians outside the company get a real look.
The Number Nobody Can Agree On: 719 vs. 722
Here is the discrepancy in plain terms. As of today, OpenAI’s GitHub repository lists 719 manuscripts organized into 372 families. But the initial wave of coverage around the October 6 release described 722 manuscripts, a number that has since been repeated across multiple reports. Three manuscripts is not a rounding error in a field built on exact counting, and nobody involved has published a reconciliation explaining whether papers were pulled, merged, or simply miscounted in the first pass.
It is worth sitting with how unusual this is. A software company shipping a slightly wrong changelog number is routine. A math lab publishing OpenAI math papers and getting its own manuscript count wrong, in the same week it is asking the field to trust its rigor, is a different kind of problem. Until OpenAI clarifies which figure is authoritative, both 719 and 722 should be treated as provisional.
Why a Head-Count Error Is a Credibility Problem, Not a Footnote
Mathematics runs on precision. A proof either closes every gap in an argument or it doesn’t, and a result either holds under every edge case or it gets thrown out. Against that backdrop, a lab that cannot settle on how many papers it just published invites a fair question: if the bookkeeping is shaky, how much scrutiny did the actual mathematics get before publication? That question is exactly why this release has become a flashpoint rather than a quiet GitHub commit.
The discrepancy also complicates independent verification. Reviewers trying to audit the collection need a stable baseline to work against. A moving manuscript count means outside mathematicians can’t even agree with OpenAI on what the full set looks like, before they’ve opened a single file to check the logic inside it.
Lean Proofs Don’t Mean “Verified”
A chunk of the repository’s credibility rests on Lean, the proof assistant and programming language that lets a computer mechanically check whether a formal proof is logically sound. Including Lean formalizations for many of the results is a real step toward accountability, since it gives outside reviewers a machine-checkable artifact instead of just prose and trust. But it is easy to overstate what this buys OpenAI. A simplified illustration of how a Lean statement is structured looks something like this:
theorem example_result (n : Nat) (h : n > 0) :
n + 1 > n := by
exact Nat.lt_succ_self nLean can confirm that a chain of logical steps follows the rules once someone has translated an informal proof into its strict syntax. It cannot confirm that the informal statement being formalized is interesting, novel, or correctly describes the problem the paper claims to solve. Having Lean files in a repository full of OpenAI math papers is a meaningful credibility signal. It is not the same thing as peer review, and OpenAI has not claimed otherwise.
Enter AGMAI, the Group Vetting the Vetters
OpenAI tied this release to AGMAI, described in the announcement as an independent advisory group of mathematicians. The company has not publicly detailed exactly what AGMAI reviewed or signed off on before publication, and the available reporting does not establish a full account of the group’s process. What is clear is that OpenAI felt it needed outside mathematical cover for this release, which is itself a signal of how sensitive the company understands this moment to be. Publishing unverified proofs under your own name is one thing. Publishing them with an advisory group’s name attached raises the stakes if the claims don’t hold up.
The mathematical community has not independently confirmed or peer-reviewed the claims in the collection, and that gap between “advisory group involved” and “peer reviewed” is doing a lot of quiet work in how this story should be read.
The Model Nobody Has Seen
OpenAI has not named the model behind this work, and the model itself has not been released to the public in any form. That is an odd combination for a company that typically ties its biggest claims to a named, shippable product. Readers get the output of a system they cannot test, interrogate, or compare against a public benchmark, which means every claim in the repository has to be taken on the strength of OpenAI’s own framing until someone outside the company reproduces the results independently.
That pattern puts this release in a different category from a typical model launch like GPT-6’s rollout into ChatGPT, where the product itself is available for anyone to probe. Here, the proof artifacts are public. The thing that generated them is not.
A Year of Escalating Math Claims: How We Got Here
This week’s GitHub drop is the latest entry in a pattern, not an isolated event. OpenAI has moved through several distinct disclosures this year, each adding a new number to track and a new round of scrutiny to absorb.
| Disclosure | Headline Number | Verification Status |
|---|---|---|
| Initial manuscript drop (widely reported) | 722 manuscripts | Not independently confirmed |
| Funded workshops to vet proofs | 372 proofs funded for review | Workshops described as in progress |
| Verification-fight coverage | 377 claims disputed | Specialist review still required |
| Current GitHub repository (Oct. 6-8) | 719 manuscripts / 372 families | Count discrepancy with 722 unresolved |
Each of those numbers, 722, 372, and 377, has circulated in separate rounds of coverage, including an earlier report on the 722-manuscript drop, OpenAI’s move to fund workshops vetting 372 proofs, and the subsequent fight over 377 disputed claims. Treating those figures as interchangeable is a mistake plenty of casual coverage has already made this week. They describe different stages of the same unfolding story, not three versions of one fact.
Competitive Landscape: OpenAI Isn’t the First Lab Here
AI labs chasing math credibility is not new. Google DeepMind’s AlphaProof and AlphaGeometry systems drew global attention in 2024 after reaching medal-level performance on International Mathematical Olympiad problems, a result the company published alongside its own formal-verification tooling built on Lean. That earlier push set the template OpenAI is now following: pair an AI math claim with a machine-checkable proof format, lean on the credibility of formal verification tools the mathematics community already trusts, and let the Lean community’s existing infrastructure do some of the reputational heavy lifting.
Where this release diverges from that template is scale and opacity. DeepMind’s Olympiad results were tied to a defined, public competition with known-correct answers to check against. OpenAI’s math manuscripts cover open-ended problems across multiple fields, generated by a model nobody outside the company has touched, with a headline manuscript count the company itself can’t currently confirm. That is a harder claim to verify and a much easier one to doubt.
What’s Confirmed vs. What’s Still Unconfirmed
Separating what reporting has actually nailed down from what is circulating as rumor matters more than usual here, given how much speculation has attached itself to this release.
| Claim | Status |
|---|---|
| Manuscripts published on GitHub Oct. 6, 2026 | Confirmed |
| Repository currently lists 719 manuscripts in 372 families | Confirmed (current listing) |
| Earlier reports describing 722 manuscripts | Confirmed as reported, unreconciled with 719 |
| Model evaluated against roughly 4,000 problems | Confirmed |
| Many proofs include Lean formalizations | Confirmed |
| AGMAI involved as independent mathematical advisory group | Confirmed, scope of review not detailed |
| Model publicly named or released | Not done |
| Collection contains exactly 377 new mathematical results | Unconfirmed |
| Papers span exactly 17 areas of mathematics | Unconfirmed |
| Model solved the Navier-Stokes problem | Unconfirmed |
| Manuscripts independently peer-reviewed | Not yet, per available reporting |
That table is the real story under the headline. A lab is claiming a mathematical milestone while conceding, in its own language, that the results sit at different stages of verification. Readers chasing a clean “AI solves math” narrative should notice how much of that table still reads unconfirmed.
The Verification Bottleneck Nobody Talks About
Even with Lean files attached, checking 700-plus manuscripts properly is a slow, human-intensive process. Formal verification confirms logical steps follow the rules of the proof system, but someone still has to confirm the formalization matches the intended mathematical claim, that the claim is actually new, and that it is not a restatement of existing work dressed up differently. That kind of review is exactly what graduate students and specialist referees spend months doing for a single paper in a peer-reviewed journal, not something a crowd of volunteers resolves in a weekend because a GitHub repository went public.
That bottleneck is the practical reason this story will not resolve quickly. Mathematicians can skim a repository in an afternoon. Actually confirming or refuting hundreds of individual results takes the kind of sustained specialist attention that academic math simply does not have lying around in surplus.
Historical Context: Computers and Math Have Fought Before
This is not the first time mathematicians have argued about whether a machine-assisted proof counts as real mathematics. The four-color theorem, proved with heavy computer assistance in 1976, spent years facing skepticism from mathematicians who objected to verifying a proof too large for any human to check by hand. It took decades, and later a fully formal verification, before that computer-assisted result became fully accepted. The pattern repeating with large language models is the same argument in new clothing: a claim arrives faster than the community’s capacity to check it, and trust has to be rebuilt the slow way, result by result, rather than granted upfront because the headline number is large.
Market and Industry Impact
For OpenAI, the stakes extend past mathematics. The company has leaned on headline AI capability claims to justify its valuation and its pace of model releases, a pace that has already drawn internal friction, including the departure of safety staff covered in earlier reporting on OpenAI’s safety lead quitting after a string of twelve model launches. A disputed manuscript count on a flagship math release feeds directly into the broader question of whether OpenAI’s announcement pace is outrunning its own verification processes.
It also matters for how rival labs calibrate their own announcements. If OpenAI absorbs real reputational cost for publishing claims that are, by its own admission, only partially verified, competitors get a data point for how far they can push unverified capability claims before the market and the research community push back. Executives on both sides of the AI-risk debate, including comments captured in coverage of Altman’s split with Anthropic over how to handle AI risk, have already staked out different tolerances for exactly this kind of move-fast-and-clean-up-later release strategy.
Predictions: Where This Story Goes Next
- OpenAI reconciles the 719 vs. 722 manuscript count within days, under pressure from the attention the discrepancy has drawn this week.
- Full independent verification of the collection stretches into months, not days, simply because the volume of material outpaces available specialist reviewer time.
- At least one widely publicized result gets quietly revised or withdrawn once outside mathematicians work through the backlog, consistent with OpenAI’s own framing that results sit at different verification stages.
- Rival labs respond with their own math-focused disclosures rather than ceding the “AI for mathematics” narrative to OpenAI, continuing the pattern set by DeepMind’s earlier Olympiad work.
- Pressure builds on OpenAI to eventually name the model publicly, especially if AGMAI or outside mathematicians push for more transparency about what generated the results.
What Readers and Researchers Should Watch For
Anyone trying to follow this story cleanly should track three things rather than the headline count alone. Watch whether OpenAI publishes a reconciled, final manuscript number. Watch whether any independent mathematician or institution issues a specific confirmation or refutation of individual results, rather than a general comment. And watch whether OpenAI eventually names the model that produced the work. Those three data points will tell you more about whether this release holds up than any restatement of the 719, 722, 372, or 377 figures currently circulating.
Frequently Asked Questions
How many OpenAI math papers were actually published?
As of October 8, 2026, OpenAI’s GitHub repository lists 719 manuscripts organized into 372 result families. Earlier reporting around the October 6 release described 722 manuscripts, and that discrepancy has not been publicly resolved.
What model produced these mathematics manuscripts?
OpenAI has not named the model. It is described only as an unreleased internal model, and it has not been made available to the public in any form.
Does including Lean proofs mean the results are verified?
No. Lean formalizations let reviewers mechanically check that a formal proof’s logical steps are valid, but that is different from confirming the underlying mathematical claim is novel, correctly stated, or independently peer-reviewed. OpenAI itself has said the published results sit at different stages of verification.
What is AGMAI and what role did it play?
AGMAI is described in the release as an independent advisory group of mathematicians that OpenAI associated with this disclosure. The full scope of what AGMAI reviewed has not been detailed in available reporting, and AGMAI’s involvement is not the same as formal peer review by the broader mathematics community.
Did the model solve the Navier-Stokes problem?
That claim is unconfirmed. It has circulated in discussion of the release, but available sourcing does not establish it as a verified fact, and readers should treat it as unresolved pending specialist review.
How does this compare to OpenAI’s earlier 722-manuscript and 377-claim stories?
Those are different stages of the same ongoing disclosure, not duplicate figures. The 722 figure referred to the initial reported drop, 372 referred to proofs funded for workshop-based review, 377 referred to disputed claims covered in a separate verification-fight story, and the current GitHub listing shows 719 manuscripts in 372 families today.
Has any other AI lab published similar mathematics work?
Yes. Google DeepMind’s AlphaProof and AlphaGeometry systems reached medal-level performance on International Mathematical Olympiad problems in 2024, also using Lean-based formal verification, though against a defined competition format rather than an open-ended manuscript collection.
When will these manuscripts be independently confirmed?
No timeline has been given. Given the volume of material and the specialist review each result requires, independent verification of a collection this size realistically takes months rather than days.
