ARC-AGI v2: leaderboard

120 public grid puzzles (plus 1,000 training tasks) from the ARC Prize Foundation (2025) where a model infers a transformation from a few examples; pass@2, human panel 60%, o3-preview-low 4%.

Source: arcprize.org.

Interactive version: theaggregate.ai/benchmark?slug=arc-agi-v2 · How It Works · Data refreshed daily, snapshot 2026-09-05.