OpenAI Model Astra Tackles 10 Long-Unsolved Math Problems With Lean 4 Proofs
OpenAI is now grading its models on unsolved research questions rather than exam problems, publishing proofs a computer can verify.
Reporting from 1 source: GIGAZINE.
OpenAI announced on August 1, 2026, that an internal version of its next flagship model, Astra, produced new results on 10 unsolved problems in mathematics and theoretical computer science. The targets had no major progress for at least a decade. OpenAI published a 249-page paper and machine-checkable proof data, with arguments formalized in Lean 4 for independent verification.
OpenAI has moved beyond exam benchmarks. An internal version of Astra, its next flagship model, worked on problems that had seen no major progress for at least 10 years, and the company published all 10 results in a 249-page paper with machine-checkable proof data.
The mathematical arguments were drafted by human staff into paper manuscripts, then Astra formalized each one in Lean 4, software that verifies each reasoning step. The GitHub repository includes Lean-format proof data and tools for independent verification.
The results include new upper bounds for sphere packing densities, a counterexample to Connes's rigidity conjecture, and new lower bounds for arithmetic circuits computing the permanent. For binary codes and spherical codes, general upper bounds improved for the first time since 1977 and 1978.
- Sphere packing densities: Improved upper bounds in high dimensions to the limits obtained by the Cohn-Elkies method.
- Binary and spherical codes: Exponentially stronger upper bounds for maximum size, first general improvement since 1977 and 1978.
- Non-sofic group: Constructed a group that cannot be approximated by finite permutations.
- Connes's rigidity conjecture: A counterexample where a group is not uniquely determined by its von Neumann algebra.
- Arithmetic circuit complexity: New lower bounds for computing the permanent of a matrix.
- Parallel repetition theorem: Extended to any finite two-player game using quantum entanglement.
- Closest vector problem: Approximation is hard even with polynomial factor errors relative to lattice dimension.
- Ehrhart's volume conjecture: Proved for convex bodies where the centroid is the only interior lattice point.
- Multicolor triangle Ramsey numbers: Superexponential lower bounds proved.
- Extremal graph theory: Counterexamples to a compactness conjecture and a degeneracy conjecture.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.