OpenAI released 719 math manuscripts, with Lean covering 42% of top line results

OpenAI has opened a repository containing 719 mathematical manuscripts grouped into 372 families. Its README says the vast majority were produced with one procedure: an unreleased internal model was given roughly 4,000 open problems and used an average of three hours of ChatGPT Pro thinking compute for each result found.

One model turned open problems into a publication batch

The repository contains PDFs, source files, citation data and supporting proof artifacts. OpenAI says some outputs build on earlier model results and that the materials sit at different stages of verification. It also released abridged reasoning summaries for ten selected families.

The initial batch contained 722 manuscripts. The current catalogue lists 719 after three were withdrawn. In his review, Zvi Mowshowitz emphasizes the breadth and says the set includes 90 problems from the top 500 in the Proof of Atlas ranking. That is an external assessment of significance, not OpenAI's own classification.

The scarce resource is no longer the first proof draft

If one model can produce hundreds of candidate results, mathematical labor shifts toward verification, explanation and integration into the field. A manuscript may contain a correct idea, but scientific value arrives when experts understand its assumptions, compare it with the literature and can build on it.

That changes research economics. Generation can scale with compute, while careful reading still consumes the time of scarce specialists. A mass release therefore creates more than knowledge. It also creates a work queue for a community that does not own the model and cannot question its author.

Lean checks a formal statement, not the whole scientific story

OpenAI reports Lean formalizations for about 42% of top line results and acknowledges that unformalized work may contain issues. Three withdrawn manuscripts shortly after release show why the gap between a candidate proof and an accepted result matters.

Even a valid Lean formalization does not settle every question. Reviewers must establish that the formal statement matches the known open problem, that important conditions are present and that the result fits existing mathematics. AGMAI therefore described publication as the beginning of human understanding, not its completion.

The pace of corrections will show whether this works as science

More useful than the PDF count will be revision history, the share of added formalizations, independent reproduction and clear exposition of the central results. OpenAI promises to preserve public versions and keep adding Lean proofs, making it possible to track which claims survive and how quickly errors are repaired.

Community participation is the second test. If mathematicians gain tools, compute and room to pose their own questions, the repository could accelerate the field. If they mostly inherit the job of checking hundreds of difficult outputs, the model has accelerated manuscript production while sending someone else the bill for understanding.

Lilith's verdict

OpenAI delivered 719 crates of manuscripts to the library. Now we find out whether they contain new maps of mathematics or mostly extra shifts for proof checkers.