On August 1, OpenAI introduced Astra, its next major model family, not with a product launch but with a research drop: an internal, unreleased version of Astra had worked through ten previously unsolved problems spanning group theory, combinatorial geometry, and theoretical computer science, including constructing an example that settles the long-open question of whether non-sofic groups exist and pushing sphere-packing density bounds closer to the Cohn-Elkies threshold. What makes this worth a developer's attention rather than filing it under "another benchmark story" is how OpenAI chose to substantiate the claim. Instead of asking people to trust a leaderboard number, they published every result as a Lean 4 formalization on GitHub — a proof written in a language a theorem-checker can mechanically verify line by line. That distinction matters enormously for anyone evaluating AI capability claims: a Lean-checked proof either compiles against the verifier or it doesn't, so there's no room for the plausible-sounding-but-wrong output that plagues natural-language AI claims. Independent mathematicians, including Fields Medalist Timothy Gowers, reviewed several of the proofs and said at least one was strong enough to recommend for a top journal without hesitation. OpenAI also disclosed the compute cost — around $2,000 total across all ten problems — which reframes a certain slice of frontier mathematical research as something closer to a cheap, parallelizable compute job than a scarce-genius bottleneck. The practical takeaway for builders is less about mathematics specifically and more about a verification pattern worth stealing: if you're shipping an AI system whose outputs need to be trusted, pairing it with a formal or mechanically-checkable verification layer (Lean, TLA+, property-based tests, SMT solvers) turns "the model says so" into "the model says so, and here's a proof object you can check yourself," which is a fundamentally stronger trust story than benchmark scores alone.