OpenAI Shares Math Results From Internal Model With Lean Proofs

OpenAI publishes math results with Lean proofs

OpenAI has published a collection of new mathematical results generated by an internal frontier model and made the accompanying Lean proof formalizations available through a GitHub repository. The release, announced on October 6, 2026, marks a notable step in how the company approaches sharing scientific discoveries produced by its most advanced systems. Rather than simply announcing breakthroughs, OpenAI has chosen to expose the underlying reasoning traces, compute estimates, and formal verification artifacts that allow the mathematics community to scrutinize and build upon the work.

The company consulted with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study while developing its release practices. That group's public recommendations, issued in late September, helped shape the protocols now in place for paper revisions, citation standards, and the structure of the GitHub repository itself. OpenAI has indicated it will continue exploring alternative community-hosted platforms that might better align with the advisory group's guidelines for future disclosures.

Each result in the repository comes with a summary of the model's reasoning process, giving researchers a window into how the system approached the problem. The company also disclosed compute requirements, noting that the average result consumed resources roughly equivalent to three hours of ChatGPT Pro usage. Statistics about the number of problems attempted versus those solved provide additional context about the model's success rate and the scope of the exploration.

Why Lean formalization matters for math AI

The decision to formalize proofs in Lean carries particular significance for both mathematicians and AI researchers. Lean's design allows a computer to mechanically verify every logical step, eliminating the ambiguity that can creep into natural language proofs. By providing these formalizations, OpenAI enables independent verification without requiring human experts to manually check each derivation. The company has committed to adding more formalizations to the repository as they become available, suggesting the current release represents a starting point rather than a complete catalog.

OpenAI funds AI mathematics workshops

Beyond the immediate artifacts, OpenAI has pledged to fund workshops, conferences, and special programs focused on understanding major mathematical results produced by AI systems. This investment acknowledges that the value of these discoveries depends on the human community's ability to absorb, extend, and teach them. The company frames this as part of a broader effort to empower scientists with state-of-the-art capabilities while maintaining responsible release practices for the models themselves.

Practical takeaways for AI developers

For developers and researchers building with AI, this release offers several practical takeaways. The transparency around compute costs provides a concrete benchmark for estimating resources needed to tackle similar problems. The reasoning summaries demonstrate how chain-of-thought outputs can be structured for scientific domains, which may inform prompt engineering and evaluation strategies. The Lean formalizations show how AI-generated proofs can be integrated into existing verification workflows, potentially accelerating the adoption of formal methods in both pure mathematics and software verification.

The emphasis on community feedback and evolving standards also signals that OpenAI views its disclosure practices as iterative. Builders who rely on frontier models for scientific work should expect the norms around result sharing, attribution, and reproducibility to continue shifting. Engaging with the GitHub repository directly, examining the reasoning traces, testing the formalizations, and participating in the planned workshops, would be the most direct way to influence those emerging standards.

Mathematicians and computer scientists who have hesitated to incorporate AI tools into their workflows now have a concrete case study showing how frontier model outputs can be packaged for rigorous scrutiny. The combination of natural language reasoning, compute transparency, and machine-checkable proofs addresses many of the objections that have kept formal verification on the margins of mainstream mathematical practice. Whether this particular model becomes widely accessible or remains internal, the release pattern establishes a template that other labs will likely follow or react against.

Readers should monitor the GitHub repository for updates, particularly the addition of new formalizations and any revisions to the papers themselves. The planned workshop series will likely announce dates and participation details in the coming months. Anyone evaluating AI systems for mathematical or scientific reasoning should treat this release as a benchmark for what comprehensive disclosure looks like, and adjust their expectations for future model evaluations accordingly.

Source: OpenAI