OpenAI Math Publishes 722 AI-Generated Manuscripts

The important story is less “AI solved mathematics” than the emergence of a scalable, inspectable workflow for generating and validating mathematical hypotheses.

• –The model was tested on roughly 4,000 problems, then distilled into 722 manuscripts—evidence of high-throughput exploration across the long tail of open research.
• –Lean formalizations and versioned citations make the release unusually useful for reproducibility, though many manuscripts still lack formal verification.
• –Each result averaged roughly three hours of ChatGPT Pro-equivalent thinking compute, so the capability currently depends on substantial inference budgets and parallel search.
• –OpenAI’s consultation with the independent AGMAI group signals that attribution, validation, and responsible disclosure are becoming core infrastructure for AI-generated research.
• –Developers can inspect the manuscripts and proof artifacts, but the model itself remains unreleased, limiting direct experimentation and replication.