tl;dr: GPT-6 Astra scored 62.7% ($26K) on ARC-AGI-3 Semi-Private using ARC's Standard harness and 99.9% ($19K) using the Provider Adapter harness, both state-of-the-art. Notably, Astra beat the median human's action efficiency on 96% of levels, using 51.7% fewer actions on average, and developed its own compact algebraic notation to model game mechanics. ARC's authors emphasize this is meaningful progress toward generalization but not proof of AGI, given the benchmark's bounded scope.