OpenAI Releases 722 Math Proofs From One Agent Prompt

OpenAI has published 722 mathematical manuscripts organized into 372 result families, all produced by an internal frontier model the company has still not released. According to a spokesperson, nearly every result came from a single prompt handed to a single AI agent, a sharp contrast with the large agent swarms behind the lab's earlier mathematics breakthroughs.

What happened

The collection now sits in a public GitHub repository, openai/math, and includes papers, supporting proof artifacts, Lean formalizations for many results, and ten abridged summaries of the model's reasoning.

OpenAI says it evaluated roughly 4,000 problems and that the average published result consumed compute equivalent to about three hours of ChatGPT Pro thinking, according to The Verge. The repository warns explicitly that the collection mixes results at different stages of verification and that some unformalized results "could have issues."

The release was expected. In September, OpenAI said the same internal model had resolved more than 100 long-standing open problems, following its Navier-Stokes announcement. This batch replaces hints with documents, complete with revision and citation protocols.

Why it matters for builders

The headline number is 372 families, but the operational detail is more interesting: a single agent, driven by a single prompt, produced almost all of them. That undercuts the assumption that frontier research work requires sprawling multi-agent orchestration. If one agent with enough thinking budget can work through thousands of open problems, the engineering premium shifts from orchestration to context management, verification, and reproducibility.

Verification is where the release is thinnest. Lean formalizations cover many but not all manuscripts, and OpenAI itself flags that unverified results may contain errors. AGMAI, the independent Advisory Group on Mathematics and AI that OpenAI set up in September, published recommendations urging labs to release results promptly, disclose model names, prompts and compute costs, and stop treating math releases as marketing vehicles. The group is now the de facto referee for how these drops get judged.

Not everyone is satisfied. WIRED reports that mathematicians describe a "perception of mobster behavior" from leading AI companies, and that OpenAI had assured attendees at an August gathering it would not publish everything at once.

What comes next

The mathematical community will need months to assess the papers. For builders, the practical signal is the release format rather than the result count: proof artifacts and machine-checkable certificates are becoming the default way frontier research claims are shipped, and any agent stack expected to work on hard problems will need to treat formal verification as a first-class step rather than an afterthought.