GPT-6.2 Astra formalizes the Smoktunowicz gap theorem
GPT-6.2 Astra Formalizes Smoktunowicz's Gap Theorem
Vals AI reports that GPT-6.2 Astra solved 49 of the 56 formalization tasks in Autoform-Hard, and is the only model in its tests to formalize Smoktunowicz's 2006 theorem on the growth of graded algebras; GPT-6.2 Sol and Claude Fable 5.2 each solved 48. Astra ran for 94 minutes without human intervention in a sandbox with Lean, Mathlib and a shell, wrote 1794 lines of Lean, and cost $21.58 at published prices. It once submitted a partial proof with a sorry gap at about 47 minutes, which was rejected, and then truthfully reported it as unresolved; because it fell short of the $20 abandonment threshold it was sent back to continue, switched to a new approach, and a complete proof was accepted 39 minutes after the unresolved report. The proof underwent clean-room scoring and human review and used only Lean's standard axioms; the authors say it is the only formalization of this theorem they know of, and the capacity argument in the latter half does not appear in the published proof.
The proof underwent clean-room scoring and human review; the latter-half route differs from the published proof, and it is a concrete sample for observing frontier models' autonomous mathematical capability.