OpenAI’s Astra Model Solves 10 Open Math Problems via Lean 4 Proofs

OpenAI has unveiled Astra, a new internal model that has resolved ten long-standing open problems in mathematics and theoretical computer science. The announcement marks a significant shift in how AI labs demonstrate reasoning capabilities, moving away from anecdotal success toward formal, machine-verifiable proof.

To support these claims, OpenAI released a 249-page manuscript dated August 2026 alongside machine-checkable Lean 4 certificates hosted on GitHub. These certificates allow the global mathematical community to independently verify the results without requiring direct access to the proprietary Astra model. Among the ten solved problems was the first explicit construction of a non-sofic group, a major breakthrough in group theory. The model also successfully disproved Connes's rigidity conjecture and resolved several problems from the renowned Paul Erdős catalogue.

The Cost of Discovery

Despite the complexity of the problems solved, the computational overhead for these specific breakthroughs was remarkably low. OpenAI reported that the total compute cost to find all ten solutions was approximately $2,000, based on Sol API rates. This figure suggests a high level of efficiency in the model's search and reasoning processes, contrasting with the massive compute clusters typically associated with training large-scale frontier models.

Bridging the Trust Gap

The decision to provide Lean 4 certificates addresses a critical trust gap between AI developers and the academic community. Traditionally, AI-generated mathematical claims have been met with skepticism due to the "black box" nature of neural networks and the tendency for models to hallucinate plausible-looking but incorrect proofs. By utilizing a formal verification language like Lean, OpenAI has shifted the burden of proof from the model's reputation to the mathematical logic itself, which can be audited by any computer running the Lean compiler.

What's Next

While the mathematical community can now verify the proofs, the full capabilities and release date of the Astra model remain unconfirmed. The industry is now watching to see if OpenAI will transition Astra from an internal research tool to a public product and whether this "proof-first" announcement style will become the new standard for validating artificial general intelligence (AGI) milestones. For now, the focus remains on the 249-page manuscript and the GitHub repository as the primary evidence of the model's reasoning power.