OpenAI dumps 722 math papers from an unreleased model

OpenAI put a GitHub repo online with 722 manuscripts of new mathematical results, grouped into 372 families. Nearly all of them came from the same procedure, run on an unreleased internal frontier model.

The numbers behind it: the model was given roughly 4,000 research problems. Each result took about three hours of ChatGPT Pro thinking compute on average. OpenAI then merged the outputs into families and kept only the ones it judged significant enough to publish. So this is a filtered catalog from a large sweep, not 4,000 solved problems.

This follows OpenAI's August release of ten proofs of long-open problems. Those came from an internal version of its next model, Astra, and shipped with machine-checkable Lean certificates. This drop is much wider and shallower: hundreds of results, with reasoning summaries and proof artifacts alongside. Verification of the full set is still up to the math community, and one of the August results already drew a refutation.

Why it matters: a few hours of compute per result is cheap enough that AI-generated math is now a volume problem, and checking it, not producing it, is the bottleneck.