OpenAI Posted 722 Math Papers. No Workshop Date.

The number people will repeat is 722. The number that tells you whether anyone outside the lab can check the work is still missing. On Oct. 6, 2026, OpenAI published a public GitHub repository of mathematical manuscripts from an internal model. Aamir Khollam at Interesting Engineering describes 722 manuscripts in 372 research families, plus Lean formalizations so a computer can check arguments. Unite.AI’s same-day account puts the repo at github.com/openai/math, under Apache-2.0. Tom Vance at Shattered, updated Oct. 7, is the one who stays on the quieter half: OpenAI says it will fund workshops so mathematicians can work through what the model produced, and it has not published a date, a host, or a dollar figure.

We wrote the model OpenAI did not ship in the Astra shelving. We wrote teams stuck in pilot in Collibra’s count of the hours. We wrote the seat-and-meter version of agents in Microsoft’s Copilot super app. This week is not another product page. It is a pile of papers and a promise about who gets to read them slowly.

722 manuscripts are not 4,000 problems

Interesting Engineering’s headline says the release tackles 4,000 problems with Lean-checked proofs. Unite.AI, citing the README, says the model was posed approximately 4,000 problems, and that the average result used the equivalent compute of roughly three hours of ChatGPT Pro thinking. The published catalogue is 722 manuscripts organized into 372 families. Those are three different counts. Treating them as one “4,000 proofs” number is how a recap gets ahead of the repo.

A family, in Unite.AI’s reading of the README, groups related papers. It may include a principal result, companion arguments, consequences, or alternative proofs, and each family is classified by discipline. A preprints directory holds PDFs, source files, and manuscripts. So 372 is not 372 theorems in the sense a department colloquium would use the word. It is 372 folders of related writeups. 722 is the manuscript count inside that structure. 4,000 is the size of the problem set the model was asked to attempt, before significance filtering and aggregation produced the catalogue. If a summary says OpenAI “solved 4,000 problems,” it has collapsed the attempt into the publication. I would not.

The README, again via Unite.AI, says OpenAI expanded these evaluations after performance on its existing mathematical evaluations saturated, and that some outputs build on earlier results from its models. Two named exceptions departed from the fixed procedure. The extract I have cuts the names off. I am not going to finish that sentence from memory. The procedure is the part that is intact: pose a large set, require a level of significance, aggregate into families and manuscripts, publish that catalogue. Exceptions exist. I cannot tell you what they were from the paragraphs in front of me.

Interesting Engineering gives two examples without claiming either one is a solved famous problem in the way a press office would want. One concerns the irrationality exponent of pi, the quantity that measures how closely rational numbers can approximate pi. The model produced a result about that approximation. Another family examines NP-hardness. Khollam’s wording is careful enough that I can repeat it: the model’s work adds mathematical results to that body of research. He does not, in the lines I am using, say a complexity theorist has signed off. Lean can check a formalization. It cannot tell you the problem was the one you think it was, or that the informal paper matches the formal proof, unless someone reads both. That reading is the workshop. The workshop has no date.

The advisory group asked for a process, not a vibe

OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. Interesting Engineering quotes the Oct. 6 post to that effect. Unite.AI and Shattered agree on the body and the host. Shattered notes the Princeton address and the Einstein and Gödel history. I do not need the campus lore. I need the document the group actually published.

Unite.AI says the group posted its responsible-release recommendations on Sept. 29, 2026, at agmai.org/general-sep29, after more than 600 replies from the mathematical community. The group says the recommendations are supported by a clear plurality of respondents. A footnote, as Unite.AI reports it, says the survey asked about a specific situation: OpenAI had announced the existence of many results without giving details. That footnote matters more than the reply count. The consultation was not a blank “how should labs behave.” It was a reaction to a prior announcement that withheld the mathematics. The Oct. 6 repo is, in part, an answer to that complaint. Whether it is a sufficient answer is what the missing workshop is supposed to test.

The document, in Unite.AI’s summary, opens by observing that some frontier labs test advanced problems on proprietary models the rest of the community cannot use. It says the group does not endorse that practice. The quote in the extract breaks off mid-sentence, so I will not complete it. The stance that survives without the missing words is already strong: do not run advanced mathematics on a model nobody else can touch, and then ask the field to trust a count.

The activities the group lists as candidates are concrete. Conferences. Summer schools. Targeted workshops. Longer-term working groups. Books or long expository articles. Decisions about which of those to support should sit with existing nonprofit institutions, not with the labs. The group also wants the mathematical community to have broad, equitable access to publicly available models. That last line is the one a lab can nod at and then miss. A public GitHub repo of manuscripts is not the same thing as access to the model that wrote them. Unite.AI’s headline calls the source an unreleased model. The papers are out. The system that produced them is not, except insofar as OpenAI points at an earlier internal result.

What “internal frontier model” points at

Unite.AI follows the phrase in OpenAI’s post. It links to the lab’s Sept. 8, 2026 announcement that an internal system had produced a Lean-formalized proof about the Navier–Stokes equations: smooth three-dimensional fluid motion can develop a singularity in finite time. OpenAI said that result resolves a problem the Clay Mathematics Institute named as one of its seven Millennium Prize Problems in 2000. I am not re-litigating that proof here. I am noting why the Oct. 6 drop cites it. The lab wants this manuscript pile read as a continuation of a system that already claimed a named problem, not as a one-off export from a product model you can call this afternoon.

That framing cuts both ways. If you trusted the September claim, the October repo is more of the same pipeline, now with hundreds of families instead of one fluid-dynamics result. If you did not trust September, a larger pile does not repair the trust. It multiplies the reading list. Lean formalizations help the second camp more than the first, because a checker does not care which blog you believed. They help only for the proofs that were actually formalized. Interesting Engineering says “many” proofs were formalized, not all of them. Unite.AI says the repo includes Lean proof formalizations and abridged summaries of the model’s reasoning. An abridged reasoning trace is not a formal proof. A PDF is not a formal proof. The object a mathematician can check without trusting the lab is the Lean file, and only where it exists and matches the claim in the manuscript.

Shattered’s comparison to DeepMind is the cleanest outside benchmark for the release shape, not for the mathematics. AlphaProof and AlphaGeometry were checked against official International Mathematical Olympiad grading in 2024, and they also used Lean. DeepMind did not release the underlying models or the manuscripts the way this repository does. So OpenAI is ahead of that particular DeepMind release on public artifacts, and behind any standard that requires the model itself. Both can be true. A workshop cannot inspect a weight file that was not published. It can inspect a Lean file that was.

OpenAI’s funding line is still a pledge

This is the sentence I would not let a recap soften. Shattered says OpenAI has stated, publicly, that it intends to fund workshops, conferences, and special programs so mathematicians outside the company can work through the results. It also says OpenAI has not published dates, host institutions, or funding figures. Interesting Engineering reports the same intention: support workshops, conferences, and special programs focused on understanding major mathematical results generated by AI, and expect feedback from mathematicians to shape later disclosures. Unite.AI adds that OpenAI is still exploring other community-hosted alternatives that meet the advisory group’s guidelines, and that it is committed to improving citations, mathematical exposition, and presentation in future releases.

“Continuing to explore” and “committed to improving” are future-tense. They are not a program. A program has a room, a month, and a budget, or at least a call for proposals with a deadline. None of the three Oct. 6 and Oct. 7 pieces prints those. Vance’s question-and-answer line is the one I would keep: the funding remains a pledge rather than a scheduled program. If a later post names the Institute, or the AMS, or a dollar amount, that post is the news. This one is the absence.

The advisory group was also specific about who should decide. Nonprofit institutions, not the lab. A lab-branded workshop that flies a dozen friendly readers to San Francisco for two days would meet the word “workshop” and miss the recommendation Unite.AI attributes to the group. I cannot know what OpenAI will fund. I can know what the group asked for, and I can know that the Oct. 6 materials do not yet show a match.

What you can check this week, and what you cannot

You can open the repository. Apache-2.0 is a real license for the files that are there. You can see whether a family you care about has a Lean formalization or only a PDF and a summary. You can read the Sept. 29 recommendations and compare them to the Oct. 6 post without taking anyone’s adjective for it. You can notice that 600 replies and a plurality are not the same thing as unanimity, and that the survey was about a prior announcement that withheld details. The repo is a response to that. Volume is not the same as the process the group listed.

You cannot, from these articles, verify a single new theorem. I did not. Interesting Engineering’s pi example and the NP-hardness family are illustrations of range, number theory and complexity and physics in the same drop, not verdicts. You cannot price the workshop. You cannot enroll in it. You cannot call the internal model and reproduce a manuscript, because the model is internal. Three hours of ChatGPT Pro-equivalent compute per result, if the README figure holds, is a cost story about how the catalogue was generated. It is not a button on a public API.

The Astra piece on this site was about a model the lab declined to ship after internal tests. This piece is about artifacts the lab did ship, from a system it still will not hand you. Those are different kinds of caution. One keeps a system off the market. The other puts the homework in public and asks the field to grade it later, on a calendar that has not been printed. I would rather have the homework than a count with no files. I would not call the grading done because the files exist.

If you cover this for a team that does not read Lean, the honest summary is short. OpenAI published 722 manuscripts in 372 families on Oct. 6, from an internal model, with some proofs formalized, under Apache-2.0. An outside group at the Institute for Advanced Study had already asked labs to stop testing advanced mathematics in private and to fund community-run reading, not lab-run theater. The reading has been promised. It has not been scheduled. Until a date exists, the repo is a source dump with a checker attached to part of it, not a verification program. Treat the 4,000 figure as the size of the attempt. Treat 722 and 372 as the size of what they chose to publish. Treat the workshop as a line item that is still empty.