OpenAI Publishes 722 AI-Generated Math Manuscripts Citing Riemann Hypothesis Progress; None Peer-Reviewed

What Happened

The numbers first: 722 manuscripts, 372 problem families, in the public GitHub repository openai/math; generated by an unreleased internal OpenAI model (not any shipped product); released on the evening of October 6 (ET); roughly three hours of compute per result (explainx.ai calls it about three hours of ChatGPT-Pro-level compute); WSJ says the batch includes advances related to three of the five remaining Millennium Prize Problems, the Riemann hypothesis among them. Rutgers math professor Alex Kontorovich wrote on X that a human producing the Riemann-related result would earn an instant Fields Medal, no questions asked.

OpenAI has made several earlier math releases. In August its internal Astra model solved ten open problems in math and theoretical computer science, Lean certificates included; on September 8 OpenAI published a 166-page Navier-Stokes paper plus a Lean formalization repository, saying it would not claim the Millennium Prize; on September 21 an official blog post confirmed an internal model (training began August 28) had resolved 100+ long-standing open problems and announced an independent math advisory group hosted at the Institute for Advanced Study. Twenty-five Fields Medalists had earlier signed a severe misalignment declaration accusing labs of front-running the field with closed models. The 722 manuscripts on the night of October 6 are the largest of these releases.

Key Facts

1. 722 papers, zero peer review: None of the 722 manuscripts has been peer-reviewed. Lean formalization covers only some of the results — explainx.ai’s October 7 follow-up checked the Lean status of eight headline results one by one, and the answer varies.
2. Related advances are not a proof of the Riemann hypothesis: WSJ’s wording is advances related to three of the five remaining Millennium Prize Problems — related advances, not solutions. Claude’s earlier Riemann zeta bound (67% of zeros on the critical line) was misreported as close to proving RH, though that result does not cover the remaining 33% of zeros. OpenAI’s own wording this time is likewise “related advances.”
3. About 3 hours per result: Roughly three hours of ChatGPT-Pro-level compute per result. AGMAI (the independent advisory group) published a release card on September 29 requiring the model name, prompts, compute cost, and a ledger of failed attempts to be public. OpenAI’s release meets deposit (public repo) but names no model and publishes no prompts or costs — two points at most on that card.
4. The dispute over release norms: Twenty-five Fields Medalists asked labs to stop testing hard problems on closed models; AGMAI’s rules are cite, rewrite, independent deposit, log cost, attach Lean, list the failures. The Hacker News thread has 91 points and 114 comments.

Context

Other AI math results in 2026: in May, DeepMind’s LLM-Lean loop cracked 9 Erdős conjectures; in August, OpenAI’s Astra solved ten; on October 3, Meta published six Muse Spark-assisted math papers; Google’s Cogentic took five theoretical computer science results. OpenAI released 722 at once, as a repository.