Definition
Family of abstraction-and-reasoning benchmarks run by arc-prize, spanning ARC-AGI-1, ARC-AGI-2, and interactive adaptation suite arc-agi-3.
Key Points
- ARC-AGI-1 / ARC-AGI-2: grid/task-style abstraction; frontier models near ceiling on older sets
- claude-opus-5 Max: 97.5% ARC-AGI-1, 90.4% ARC-AGI-2 Semi-Private (2026-07-26-arc-prize-claude-opus-5-results)
- arc-agi-3: interactive unfamiliar environments — largest remaining gap vs humans