A third-party evaluation has released the latest leaderboard for the highly challenging mathematics benchmark Autoform-Hard. GPT-6.2 Astra tops the list with an accuracy of 87.50%, ahead of the tied Claude Fable 5.2 and GPT-6.2 Sol. The decisive gap comes from the 2006 Smoktunowicz gap theorem, which no one had previously been able to solve: Astra completed the world's first formal proof in just 94 minutes with zero human intervention. Because the Lean public mathematics library lacks the three major theorems on which the original proof depended, Astra hit a roadblock midway through its derivation and was about to give up—but was overruled by the system for not having reached the specified compute quota. It then defined a new mathematical tool called "capacity" on its own, bypassing the original path and carving out an entirely new proof route—without invoking or supplementing the three missing major theorems—and ultimately completed the proof. This proof method has not been seen in the relevant literature before.
