Free tools Windows power users keep installed
One-click scans. No signup required.
On October 6, 2026, OpenAI published a large public collection of AI-generated mathematics on GitHub. Its repository README currently lists 722 manuscripts grouped into 372 families, produced by an internal model. That is a count of documents. It is not a count of independently checked proofs or accepted breakthroughs. OpenAI says the manuscripts are at different stages of verification, that not all have Lean formalizations, and that “some of the unformalized results could have issues.”
The harder problem now belongs to mathematicians, not the model. They have to read, check, place and credit a volume of work that no field normally absorbs in one drop. This article covers what was released and what the numbers mean. It also explains what Lean does and does not prove, and why some mathematicians are angry. The story is moving fast, and the repository can change. Counts here are as checked on October 7, 2026.
As an Amazon Associate I earn from qualifying purchases.
What did OpenAI release?
OpenAI’s October 6 post, “Sharing AI progress in mathematics,” says the company is releasing a broad range of mathematical results from an internal frontier model. That model has not been made publicly available. The material sits in the openai/math GitHub repository, with manuscripts and supporting proof artifacts. OpenAI says many of the proofs have formalizations in Lean, with more to be added as they are obtained. It also says the repository includes protocols for paper revisions and citations.
OpenAI says it is consulting the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. It says it drew on that group’s advice and public recommendations for the release. That is OpenAI’s description of its process. The announcement does not claim that the advisory group, or anyone else, has validated the whole corpus.
#1 Best Overall
How many math papers did OpenAI publish?
The answer depends on the unit. The figures below are all OpenAI’s own, except the WIRED item.
| Figure | What it counts | Source and caveat |
|---|---|---|
| 722 | Manuscripts in the repository | OpenAI repository README, checked October 7, 2026. This is a live count and may change. |
| 372 | Families the manuscripts are grouped into | Same README. Related manuscripts are bundled, so this is not 372 independent results. |
| About 4,000 | Problems posed to the model | OpenAI README. It is an input count, not a success count. |
| Roughly three hours of ChatGPT Pro thinking | Compute used by the average result | OpenAI’s October 6 post and README. It is a company-reported compute equivalent, not a measure of human effort, cost or correctness. |
| Around 40 | Mathematicians at an August discussion convened by OpenAI | WIRED, October 6, 2026, citing attendees. |
The sources give no corpus-wide correctness rate and no exact count of Lean-formalized results. Treat any figure that implies either with suspicion.
Did AI solve hundreds of unsolved math problems?
That is not yet established. WIRED reports that attendees at the August meeting came away understanding that OpenAI’s models had solved hundreds of longstanding problems. The repository’s own framing is more careful. It offers manuscripts at varying verification stages. It does not offer a ledger of confirmed solutions to named open problems.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteSeveral things stand between “a model wrote a manuscript claiming a result” and “mathematics has a new theorem”:
- Whether the argument is correct.
- Whether the result is new, or already in the literature under another name.
- Whether the problem was really open or important.
- Whether the paper is written well enough for others to build on.
Some manuscripts may clear all four. Some may clear none. The release does not sort them for the reader.
Are OpenAI’s new math proofs verified?
Partly, and unevenly. The README says the manuscripts are at different stages of verification and that not all have accompanying Lean formalizations. It also says some unformalized results could have issues. OpenAI says it will try to fix problems quickly and preserve release history, with corrections recorded as new versions.
The earlier unit-distance result (below) was reviewed by outside experts. The sources do not show that these 722 manuscripts received the same scrutiny. The release is too recent for peer-review status to exist for most of them.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →What is Lean, and does a Lean proof mean the result is correct?
Lean is a programming language and proof assistant. A proof written in it is checked line by line by software, so a gap in the logic is rejected rather than overlooked. That makes a completed formalization strong evidence for the specific statement it encodes.
Rank #3
Two limits apply, and they are general features of formal verification, not criticisms specific to OpenAI:
- The statement has to match the claim. Lean confirms that the formal statement follows from its foundations. A human still has to confirm that the formal statement says what the paper says it proves. A subtly mistranslated definition can yield a “verified” proof of the wrong thing.
- Coverage matters. A formalization may cover a core theorem but not every claim in the manuscript. The sources do not say how fully each paper’s claims are covered.
So Lean is a strong filter, not a verdict on the whole corpus. It also says little about novelty, significance or exposition.
Why are mathematicians upset?
Reporting by WIRED (Isabella Ward, October 6) describes several separate complaints. They should not be read as one view held by the whole field, and they are not simply “mathematicians fear AI.”
Checking burden
Mathematics relies on human readers to check arguments, and that time is scarce. Hundreds of manuscripts released together, with uneven verification, push a large reviewing load onto volunteers outside the company.
Rank #4
- Exercise your mind with this collection of brainteasers, logic puzzles, and more! 359 puzzles
Release practice
According to attendees WIRED spoke to, the August meeting included discussion of how the field should respond if AI capabilities outpace human researchers. Recollections conflict about what OpenAI said regarding release timing. WIRED reports that spokesperson Lindsay McCallum was “not aware of” an assurance that results would not all be released at once. Treat that assurance as a disputed account, not a fact.
Northwestern mathematician Bryna Kra told WIRED the group wanted papers with explanations mathematicians could absorb and use, not results announced only in a blog or tweet. She said: “Math by tweet and math by press release to me is not the way to nurture the ecosystem that created the fertile ground that they have trained on.”
Credit and prior work
WIRED also describes concerns about attribution to existing literature and about the norms of academic publication. A model can reproduce or extend known ideas without pointing to where they came from. OpenAI itself says future papers should improve in citations, exposition and presentation. That is an acknowledgment that the current ones are uneven on those points.
OpenAI’s response
OpenAI disagrees with a characterization of its conduct as “mobster behavior,” McCallum told WIRED. She said the company is working with the mathematics community to navigate the future collaboratively. OpenAI says it plans workshops, conferences and programs to help people understand major AI-produced results. The sources name no partners or dates for these.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What an earlier case shows: the unit-distance counterexample
Scientific American reported on May 21, 2026 that an OpenAI model had produced a counterexample to Paul Erdős’s unit-distance conjecture, an 80-year-old question about how many pairs of points in the plane can sit at a specified distance. Mathematicians including Daniel Litt reviewed it. Timothy Gowers of the University of Cambridge said that “No previous AI-generated proof has come close” to the high standards the article describes. Litt called it “the unique interesting result produced autonomously by AI so far.” That was his assessment in May, not a ruling on the October release.
The same article stressed limits:
- The model did not prove its construction was optimal, and mathematician Will Sawin had already improved on it.
- Humans edited and interpreted the model’s output. It was not a polished, fully contextualized paper that the model published on its own.
The case shows that an AI system can find a valid and unexpected route, and that humans still decide what it means. It is a useful yardstick for the new release, and it also shows what each manuscript would need before the same description fits.
How the new release compares with earlier AI math demonstrations
No controlled evaluation compares these 722 results with earlier ones, so a capability ranking would be invented. These are the questions worth asking of any single manuscript:
| Question | Unit-distance result (May 2026) | October 2026 repository |
|---|---|---|
| Scale and unit | One result, one conjecture | 722 manuscripts in 372 families (OpenAI) |
| Type of result | Counterexample to a conjecture | Varies by manuscript; the sources do not give a breakdown |
| Formalization and review | Reviewed by outside mathematicians, per Scientific American | Uneven; some Lean formalizations, no stated corpus-wide external review |
| Exposition and citations | Humans helped interpret and present it | OpenAI says citations and exposition should improve in future papers |
| Human role | Editing and interpretation by mathematicians | Not stated for the corpus as a whole |
Can I read the papers and check the proofs myself?
You can read them. The manuscripts and supporting artifacts are public in the openai/math repository. Checking is harder. Judging an advanced proof takes expertise in the relevant subfield, and a general reader cannot reliably tell a sound argument from a flawed one. A practical approach for a non-specialist:
Quick Recap
- Open the repository README and read the verification and revision notes before any individual paper.
- Check whether the manuscript you care about has a Lean formalization. Where one exists, ask whether the formal statement matches the claim in the paper.
- Look at the version history. OpenAI says corrections are recorded as new versions, so an earlier citation may point to a superseded text.
- Wait for specialists in that area to comment before treating a result as established. Expert commentary on the unit-distance result came from named mathematicians, and similar signals will be the useful ones here.
What to watch next
- The README counts. The 722 and 372 figures are live and may change as manuscripts are revised, merged or withdrawn.
- Formalization coverage. OpenAI says more Lean formalizations will be added. An actual count would be the first concrete measure of verification.
- Independent commentary. Statements from the Institute for Advanced Study advisory group and from specialists in each manuscript’s area will matter more than the headline totals.
- Credit practice. The next batch of papers is where OpenAI’s promised improvements in citations and exposition can be judged.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
