OpenAI Drops 721 AI-Written Math Papers With Machine-Checked Proofs
OpenAI published 721 math manuscripts in 373 families from an unreleased internal model, with Lean proofs covering roughly 42% of top-line results.
Takeaways
- OpenAI released openai/math, a repo of 721 manuscripts in 373 families from an unreleased internal model.
- Roughly 42% of top-line results ship with Lean formalizations that are mechanically checked.
- Each result used about three hours of ChatGPT Pro thinking compute, drawn from ~4,000 problems.
- Ten families include abridged reasoning traces, covering Kaplansky's conjecture, Mahler conjectures, and more.
- Exceptions include a Riemann zeta zero-free region and Hodge Conjecture for CM abelian varieties.
- Mathematicians including Tao have raised verification and attribution concerns about unformalized claims.
OpenAI has published openai/math, a repository of mathematical manuscripts generated by an unreleased internal model. The collection includes PDFs, LaTeX source, Lean formalizations, compute estimates, citation blocks, and selected reasoning summaries. It carries an Apache 2.0 license and had passed 17,000 GitHub stars by the repository’s October 2026 snapshot.
The release matters because it combines research-scale mathematical claims with hundreds of machine-checked proofs. OpenAI provides the resulting artifacts, while the model, its weights, training data, and a reproducible generation pipeline remain unavailable. Researchers can audit the published work, but they cannot independently rerun the system that produced it.
OpenAI describes the manuscripts as raw research data and warns that unformalized results may contain errors. The repository also defers questions of novelty, credit, and precedence to the mathematics community, with corrections and additional formalizations promised as review proceeds.
The catalogue, file by file
The repository separates manuscripts, formal proofs, and supporting material into a compact directory structure:
preprints/contains PDFs, TeX sources, citation metadata, and build instructions for each manuscript.lean/contains formal proofs, theformalization.yamlcatalogue, and instructions for running the repository’s Comparator checks.reasoning_traces/contains abridged reasoning summaries for 10 selected result families.overview.pdfandCONTENTS.mdmap the collection by family and mathematical discipline.
A result family groups a principal claim with related arguments, consequences, or alternative proofs. Each family is tagged by discipline, allowing researchers to narrow the catalogue before opening individual manuscripts.
Why the counts differ
The public totals changed after minor revisions, and the repository uses separate units for manuscripts, families, and top-level results: