OpenAI publishes new math results from internal frontier model

Key Points

1. OpenAI released a batch of new mathematical results produced by an internal frontier model.
2. Proofs are posted on GitHub, many formalized in Lean so computers can verify them.
3. The release follows mathematician skepticism after OpenAI's Millennium Prize-related solution a month earlier.

The latest:

A new set of mathematical results generated by an internal frontier model is now public, OpenAI said, published in a GitHub repository with citation and revision protocols attached. The company said many of the proofs are formalized in Lean, a language that lets a computer check mathematical arguments. It framed the release as an effort to push the frontier of human knowledge.

Details:

• The release: OpenAI said it is publishing the results in a GitHub repository that carries protocols for paper revisions and citations, rather than a single paper. The Wall Street Journal reported the company released findings on more than 300 problems, a sharp escalation from the single high-profile solution that preceded it.
• Machine-checkable proofs: Many of the proofs are shared as formalizations in Lean, a programming language built to let proofs be verified by computer rather than by referee alone. OpenAI presented this as part of the release design, aimed at making the output testable by the mathematics community on its own terms.
• The compute disclosure: OpenAI said it is publishing estimates of compute spent, measured in terms of Pro usage on ChatGPT. The average result consumed roughly three hours of ChatGPT Pro thinking — a figure the company offered as a scale marker for what producing each result cost.
• Reasoning summaries: The repository includes 10 summaries of the model’s reasoning, alongside statistics on how many problems were attempted. OpenAI described both as steps toward scientific transparency and openness, giving outside readers a view of the attempt rate behind the published successes.
• Outside advice: OpenAI said it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study to develop best practices, and drew on the group’s advice and public recommendations in shaping how the results were released.
• Funding commitment: The company said it will fund a series of workshops, conferences and special programs devoted to understanding major results produced by AI. OpenAI did not name participating institutions or set a timeline for the first event.
• The backdrop: The release follows OpenAI’s Millennium Prize-related math solution a month earlier. The Wall Street Journal wrote that the company is trying to win back the world of math after skepticism from mathematicians, under the headline “AI Solved One Math Problem and Everyone Freaked Out. It Just Cracked Hundreds More.”
• The model: OpenAI attributed the output to an internal frontier model and did not identify it by name, say whether it will be released publicly, or specify how it differs from the systems available to ChatGPT subscribers.

Background:

Millennium Prize Problems are a set of landmark open questions in mathematics. OpenAI’s work tied to that family of problems a month before this release drew wide attention and also pushback from mathematicians over how AI-generated results should be presented and verified.

Between the lines:

The release is built around verifiability rather than announcement. Lean formalizations, compute estimates, reasoning summaries and attempt statistics all answer the specific objection mathematicians raised after the earlier Millennium Prize-related solution: that a headline claim is not a checkable proof. Consulting the Institute for Advanced Study advisory group and funding workshops point the same direction — toward acceptance by the field, not just publicity.

What’s next

Watch whether mathematicians independently verify the Lean formalizations and how many of the more than 300 problems survive scrutiny, plus the first funded workshop and any disclosure identifying the internal frontier model.