ARC-AGI-3 benchmark resets AGI progress

ARC-AGI-3 is a brutal "vibe check" for LLMs, demonstrating that scale alone hasn't bridged the gap to human-level reasoning. Frontier models like Gemini 3.1 and GPT-5.4 are effectively failing version 3's interactive tasks, even as they approach 80%+ scores on version 2. The new RHAE metric (Relative Human Action Efficiency) heavily penalizes the trial-and-error approach current models use compared to human intuition. While critics argue the tasks are biased toward human mental models, the $850k 2026 prize pool signals the industry is taking this benchmark as the definitive AGI moat. The move from static grids to 150+ hand-crafted "game" environments forces a shift toward agents that can learn without task-specific prompt engineering.