OpenAI’s unreleased Astra reportedly solved 10 open math problems, with Lean proofs attached

Don't Worry About the Vase (Zvi) · rss · 2026-08-04

OpenAI’s unreleased model Astra reportedly solved 10 major open mathematics problems in an internal test.

According to the post, OpenAI said the results came from an internal version of Astra, with humans turning the model’s arguments into manuscripts and the model then formalizing them in Lean. The solved problems span multiple fields, including sphere packing, binary and spherical codes, non-sofic groups, Connes’s rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, closest vector problem hardness, Ehrhart’s volume conjecture, multicolor Ramsey numbers, and extremal graph theory. OpenAI also said the token cost to find these solutions would have been about $2,000 at Sol API rates.

The surrounding commentary argues that this may mark a major jump in scientific reasoning and suggests the bottleneck is increasingly about asking the right question and giving the model enough test-time compute. It also notes that the results still need scrutiny, since Lean proofs do not automatically guarantee every claimed theorem is correct.

Original post →

More from Models

Models channel →