OpenAI published a post on 6 October 2026 titled “Sharing AI progress in mathematics.” It says the company is “releasing a broad range of new mathematical results produced by an internal frontier model,” and that it is publishing them in a GitHub repository, openai/math.
The repository’s README says the catalogue holds 722 manuscripts in 372 families, that the model was posed approximately 4,000 problems, and that each result used, on average, three hours of ChatGPT Pro thinking compute. These are OpenAI’s figures, and we have not checked any of the mathematics. The README also says some of the unformalized results “could have issues.”
Business Pill · THE COST OF CHECKING
A one-minute explainer of the idea behind part of this story: producing something can be cheap while checking it is not, known as verification cost. It teaches the concept, not this story’s figures.
The key insight: The release counts effort one way and checking another. The README states effort as compute, an average of three hours of ChatGPT Pro thinking per result, and states checking as partial: results “at different stages of verification,” and not all with Lean formalizations. As we read it, the figures the post and README give describe how the manuscripts were produced; the documents we read give no count of manuscripts checked by anyone outside OpenAI.
The post says the results are published in a GitHub repository “with protocols for paper revisions and citations,” and that OpenAI is “continuing to explore other community-hosted alternatives for this release which meet the committee’s guidelines.” It says the repository includes formalizations of many of the proofs in Lean, “a programming language that allows mathematical proofs to be checked by a computer,” and that more will be added as OpenAI obtains them.
It says OpenAI is also publishing “10 summaries of the model’s reasoning, estimations of compute spent in terms of Pro usage on ChatGPT, and statistics about the number of attempted problems.” It adds: “The average result used the equivalent compute of roughly three hours of ChatGPT Pro thinking.”
The post says OpenAI will be funding “a series of workshops, conferences, and special programs around the understanding of major results produced by AI,” and that it is “working to responsibly release the model that produced these results.” It describes the model only as an “internal frontier model” and names no model in this post as we read it.
How much of OpenAI’s math release links to Lean. The totals (372 families, 722 manuscripts) are OpenAI’s. The Lean counts are our own, from the repository’s CONTENTS.md and lean/formalization.yaml on 6 October 2026; the two panels use different units, families and papers, and were not reconciled.
What the Repository README Says
The README says the repository “contains mathematical manuscripts and supporting proof artifacts produced by an internal OpenAI model.” It says: “As part of model development, we evaluate our models on open research problems,” that OpenAI expanded these evaluations “after performance on our existing mathematical evaluations saturated,” and that “Some outputs build upon earlier results produced by the models.”
On size, the README says “The current catalogue contains 722 manuscripts organized into 372 families.” A family groups related papers, and each family is classified by mathematical discipline. On method, it says “The vast majority of results were obtained with the same procedure using an unreleased internal OpenAI model.” It says each result used, on average, three hours of ChatGPT Pro thinking compute with that model, and that the model was posed approximately 4,000 problems. Aggregating the output into families and manuscripts and “requiring an appropriate level of significance” produced the catalogue.
The README names two exceptions to the fixed procedure: work on a zero-free region for the Riemann zeta function and the proof of the Hodge conjecture for CM abelian varieties. It adds that the write-up for the Re(s) > 11/12 zero-free region was “human edited for readability.” It does not say, as we read it, what the different procedure was for the exceptions.
What the Collection Says It Proves
The repository’s CONTENTS.md file describes each family. We read those descriptions, not the manuscripts. They are OpenAI’s statements about what the manuscripts claim.
Family 017 is titled “The irrationality exponent of π is 2” and is described as “Proves that the irrationality exponent of π is exactly 2.” Family 087 is described as “Resolves the symmetric and nonsymmetric geometric Mahler conjectures in every dimension.” Family 196 is titled “A counterexample to Kaplansky’s zero-divisor conjecture.” Family 287 is described as “Resolves the free group factor isomorphism problem.”
Family 003 is titled “The quasi-Riemann hypothesis” and is described as resolving it. Family 032 is described as “Proves the rational Hodge conjecture for every complex CM abelian variety, in every dimension and codimension.”
These are descriptions in the repository. We did not read any of the 722 manuscripts, and we cannot say whether any of these results holds.
How Much Is Formally Checked
The README says: “This collection includes results at different stages of verification. Not all have accompanying Lean formalizations.” It also says: “Some of the unformalized results could have issues. We will endeavor to fix any such issues quickly.” It says corrections and revisions “will be recorded as new versions, with previously released versions remaining accessible.”
Our counts from the repository files: CONTENTS.md links a Lean page for 235 of the 372 families, and the formalization catalogue, lean/formalization.yaml, lists 162 papers with a formalized main result. The two counts use different units, families and papers, and we did not reconcile them.
That catalogue file has metadata fields. As we read them, its status scope is “Partial progress.”, its automation method is “agent” and its review status is “unchecked.” We do not know who set those fields or what they certify.
The Advisory Group the Post Cites
The post says OpenAI consulted the “independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study,” and drew on its advice and public recommendations. The group’s page, “Responsible Release of AI-Generated Mathematics,” is dated 29 September 2026. It says the group received over 600 replies from the mathematical community.
The page says some frontier labs are testing advanced mathematical problems on proprietary models, and states: “we do not endorse this practice, and we ask them to stop testing advanced mathematical problems on proprietary models.” It does not name a lab, as we read it.
Its recommendations include depositing results in repositories that “should not be controlled by any AI lab.” They also include that each result should come with “the name of the model, the prompts used, a (summarized) chain of thought, the time taken, and the estimated cost of computation.” Proofs should be formalized “As far as possible,” and when many results are released at once, a document should explain how many other problems of comparable difficulty the models tried and failed to solve.
The page also says: “We strongly recommend that AI labs refrain from treating the release of mathematical results as marketing vehicles to promote their models.”
Our reading, set against OpenAI’s documents: the post and README give a compute estimate, a problem count, 10 reasoning summaries and partial Lean formalization. The repository sits in OpenAI’s own GitHub organization, and the post says OpenAI is exploring community-hosted alternatives. The README calls the model an unreleased internal one and names no model. We saw no list of prompts in the post or the README.
We did not open the manuscripts or the reasoning summaries, so we cannot say what else the repository holds, and we read no statement from the group about this release.
A Related OpenAI Post
The post links the phrase “internal frontier model” to a separate OpenAI post, “On the Navier–Stokes Millennium Prize Problem.” That post says OpenAI is “sharing a solution to the Navier–Stokes existence and smoothness problem, one of the Millennium Prize Problems,” and that to solve it OpenAI “used an internal model that is significantly more capable than GPT-6 Astra.” We read that post but not its proof or its Lean files.
It reports compute in a different unit: across all attempted problems in that effort, it says, the agents sent 4.9 million messages and used about 300 billion output tokens. That is a separate effort described in a separate post. Neither document, as we read them, names the model behind the 722 manuscripts, and we do not compare the two sets of figures.
The Structural Read
The post and README describe a pipeline more than a paper. A model is posed problems, its output is aggregated into families and manuscripts, a significance threshold is applied, and the results are published with reasoning summaries and compute estimates. In the README as we read it, the figures are about 4,000 problems posed and 722 manuscripts in 372 families, and the threshold is stated only as “an appropriate level of significance.”
The check these documents describe is Lean formalization, in which a computer checks the steps of a proof. Our counts show how partial that cover is: CONTENTS.md links a Lean page for 235 of the 372 families, and the formalization catalogue lists 162 papers with a formalized main result, against 722 manuscripts. The two counts use different units and we did not reconcile them. The catalogue file’s own review field reads “unchecked.”
The advisory group the post cites asks for several things, and the documents show a mixed picture. The post and README give a compute estimate, a problem count, 10 reasoning summaries and partial Lean formalization. The repository sits in OpenAI’s own GitHub organization, where the group’s page says repositories “should not be controlled by any AI lab,” and the post says OpenAI is exploring community-hosted alternatives. The group’s page also says it does not endorse labs testing advanced problems on proprietary models. We read no statement from the group about this release.
OpenAI, in the README of its mathematics repository
“Some of the unformalized results could have issues. We will endeavor to fix any such issues quickly.”
Three Implications
EFFORT IS DISCLOSED AS COMPUTE; CHECKING IS PARTIAL The README gives an average of three hours of ChatGPT Pro thinking per result and says the collection includes results at different stages of verification. It says not all have Lean formalizations.
THE RELEASE FOLLOWS SOME OF THE ADVISORY GROUP’S ASKS, AS WE READ IT The post says OpenAI drew on the group’s public recommendations, and it gives compute estimates, problem statistics and reasoning summaries. The repository is on OpenAI’s own GitHub organization, and the post says OpenAI is exploring community-hosted alternatives.
THE MODEL IS NOT AVAILABLE TO TEST The README calls it “an unreleased internal OpenAI model,” and the post says OpenAI is “working to responsibly release the model.” The documents we read give no release date.
What Is Not Established
We did not read any of the 722 manuscripts, the overview PDF, the 10 reasoning summaries or any Lean file, and we ran no Lean check. We cannot say whether any result is correct or new.
In the documents we read, OpenAI does not name the model, does not say how the significance threshold was applied beyond “an appropriate level of significance,” and does not say who outside OpenAI, if anyone, has checked the manuscripts. The README says each “result” used three hours on average but does not say whether a result means a manuscript or a family, so we have not multiplied the average.
We read OpenAI’s post through a text copy, because the site did not return the page to our direct request. We read the repository’s README and files through GitHub’s public interfaces on 6 October 2026; GitHub showed one commit, “Initial commit,” at 21:58 UTC. We read no outside comment from mathematicians, and we did not contact OpenAI or the advisory group.
The Map of AI places more than 200 companies across nine layers of the stack, from silicon to application. It shows where model labs sit relative to the infrastructure and software companies around them.
On 6 October 2026 OpenAI said it was releasing 722 manuscripts in 372 families produced by an internal model it has not released, from about 4,000 problems posed, at an average of three hours of ChatGPT Pro thinking compute per result. The figures are OpenAI’s. Its own README says the collection includes results at different stages of verification and that some unformalized results could have issues.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece rests on OpenAI’s post of 6 October 2026, “Sharing AI progress in mathematics,” read in full as a text copy because the site did not return the page to our direct request; on the README, CONTENTS.md and lean/formalization.yaml of OpenAI’s openai/math repository, read through GitHub’s public interfaces on 6 October 2026; on the Advisory Group on Mathematics and Artificial Intelligence’s page of 29 September 2026; and on OpenAI’s linked Navier–Stokes post.
We did not read any of the 722 manuscripts, the overview PDF, the reasoning summaries or any Lean file. The counts of families with a Lean page and of papers in the formalization catalogue are our own. We haven’t checked OpenAI’s figures or any mathematics independently, and we read no outside comment from mathematicians. We did not contact OpenAI or the advisory group. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
