Frontier snapshot · August 2026
Each dot is a real task. The diagonal is parity with people. Points above it are where machines already beat the human baseline. Points below it are where people still lead — often by a lot.
AI score
Human score
Reasoning
Human
100
Untrained people on calibrated games
Frontier AI
30
Claude Opus 5 (most models <8%)
Gap -70 points on a 0–100 task scale
Give a person a new game with no manual and they learn it. Give a frontier model the same world and it mostly flails. This is the cleanest remaining gap in fluid skill-acquisition.
Interactive, no language crutches, scored against human action efficiency. The public board is still sparse.
Try an item from this testOne item from every benchmark on the map. Puzzles are grids and boards. Official items are marked; the rest are faithful to how the real test is taken, because the held-out sets cannot be published.
0 of 30 tried
Knowledge
Math
Code
Reasoning
Games
Perception
Science
Language & art
Physical world
ARC-AGI-3 scores how efficiently a player acquires a new game. People do this. Models mostly flail.
You are the pale square. Reach the marked cell. Click an adjacent tile to step. That is all you are told — same as the real test.
Adjacent tiles only.
Tap a row to pin it on the chart.