François Chollet

François Chollet (@fchollet) posted ARC Prize’s evaluation of **GPT-6 Astra** on **ARC-AGI-3**, calling it a step-function change for interactive reasoning. - **Standard harness** (max): **62.7%** on ARC-AGI-3 Semi-Private for about **$26K**. - **Provider Adapter harness** (high): **99.9%** for about **$19K** (preserves opaque reasoning state and uses compaction). - Action efficiency: Astra (max, Provider Adapter) used fewer actions than the human baseline on **96%** of levels. - Observed on-the-fly **symbolic world modeling** and a compact, game-specific DSL / algebraic shorthand; Chollet notes these behaviors used to require sophisticated harnesses and are shifting into the model. ARC Prize says saturating ARC-AGI-3 is not a claim of AGI. Full write-up by Greg Kamradt.