OpenAI publishes 722 mathematical manuscripts from an unreleased internal model across 372 research families, with explicit verification caveats

What changed

OpenAI created and published the `openai/math` repository on October 6 with 722 mathematical manuscripts grouped into 372 families, produced by an unreleased internal model. The README says the evaluation posed roughly 4,000 open research problems and that most results used a common procedure averaging about three hours of ChatGPT Pro thinking compute per result. OpenAI explicitly notes that verification is uneven, not every manuscript has a Lean formalization, and some unformalized results may contain issues.