GPT-6.2 Astra formalizes Smoktunowicz's gap theorem on Autoform-Hard
GPT-6.2 Astra Formalizes Smoktunowicz's Gap Theorem
GPT-6.2 Astra is the only model on the Autoform-Hard leaderboard to formalize Smoktunowicz's 2006 theorem on the growth of graded algebras, solving a 49th task that GPT-6.2 Sol and Claude Fable 5.2 do not. The run took 94 minutes, 146 agent turns and 1,794 lines of Lean, costing $21.58 at Astra's published price inside a sandbox with Lean, Mathlib and a shell and no human involvement. Vals.ai says the proof builds with no sorry and only standard axioms, and that after a shared opening it diverges from the published argument, replacing three deep results absent from Mathlib with a capacity argument related to Zhang's lower transcendence degree.
A benchmark write-up documents the one Autoform-Hard task that separates GPT-6.2 Astra from GPT-6.2 Sol and Claude Fable 5.2, with a line-by-line Lean proof and a route that departs from the published argument.
Source:Hacker News · AI · vals.ai