A living survey published in March 2026 analyzes 82 approaches across three versions of the Abstraction and Reasoning Corpus (ARC-AGI) benchmark. The study reveals that while top models like Opus 4.6 achieve 93.0% accuracy on ARC-AGI-1, performance drops to 68.8% on ARC-AGI-2 and 13% on ARC-AGI-3.

  • Performance degradation of 2-3x is consistent across program synthesis, neuro-symbolic, and neural paradigms from version 1 to 2.
  • Test costs fell 390x in one year, dropping from $4,500/task for o3 to $12/task for GPT-5.2, though this reflects reduced test-time parallelism.
  • Kaggle-constrained entries with 660M-8B parameters achieve competitive results, supporting the thesis that intelligence is skill-acquisition efficiency.
  • ARC Prize 2025 winners required hundreds of thousands of synthetic examples to reach only 24% on ARC-AGI-2.

The authors conclude that compositional reasoning and interactive learning remain unsolved challenges, with current systems remaining knowledge-bound rather than demonstrating true fluid intelligence.