The company released a catalogue of 722 manuscripts grouped into 372 families of results produced by an internal frontier model. OpenAI says some work has yet to undergo formal verification.
OpenAI said the model was evaluated on roughly 4,000 open mathematical problems. The company then grouped related results and selected work it considered significant enough to publish. The catalogue spans several fields, including algebra, topology, number theory, theoretical computer science and logic.
The catalogue contains 722 manuscripts arranged in 372 families of related results. OpenAI has not specified the exact number of individual problems solved. A family may include a main result, consequences, supporting arguments or alternative proofs, so the manuscript count does not equal the number of separate problems.
OpenAI said some findings remain unverified and that not all the work has been formalized in Lean. The company warned that unverified results could contain errors. Independent specialists must now examine the material. Posting the manuscripts does not mean that the mathematical community has confirmed every proof.
The new collection expands on the company’s earlier announcement about results on open mathematical problems. Previous OpenAI claims, including one involving a solution to the Navier-Stokes equation, prompted debate over AI’s role and the extent of independent verification for such proofs.