ARC-AGI-1 measures abstraction from sparse examples on visual pattern-transformation puzzles. It is used to track non-language-heavy generalization ability.
Few-shot puzzle format rewards abstraction over memorized facts; semi-private evaluation reduces direct benchmark leakage risk.
Task design captures a narrow abstraction skill rather than full assistant usefulness; low absolute pass rates make small score differences noisy.
Higher is better. Percentage of tasks solved correctly.
Max score: 100
Task design captures a narrow abstraction skill rather than full assistant usefulness.
| # | Model | Score |
|---|---|---|
| 1 | 100.0 | |
| 2 | 100.0 | |
| 3 | 99.9 | |
| 4 | 99.0 | |
| 5 | 98.5 | |
| 6 | 98.5 | |
| 7 | 98.5 | |
| 8 | Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 5 Fallback) source arc_agi Evidence observed Nov 22, 2026 | 98.5 |
| 9 | 98.5 | |
| 10 | 98.5 | |
| 11 | 98.5 | |
| 12 | 98.5 | |
| 13 | 98.5 | |
| 14 | 98.0 | |
| 15 | 98.0 | |
| 16 | 97.5 | |
| 17 | 97.5 | |
| 18 | 97.5 | |
| 19 | 97.5 | |
| 20 | 97.5 | |
| 21 | 97.5 | |
| 22 | 97.5 | |
| 23 | 97.5 | |
| 24 | 97.0 | |
| 25 | 96.5 | |
| 26 | 96.5 | |
| 27 | 96.5 | |
| 28 | 96.5 | |
| 29 | 96.3 | |
| 30 | 96.0 | |
| 31 | 95.5 | |
| 32 | 95.5 | |
| 33 | 95.5 | |
| 34 | 95.0 | |
| 35 | 95.0 | |
| 36 | 95.0 | |
| 37 | 94.9 | |
| 38 | 94.8 | |
| 39 | 94.5 | |
| 40 | 94.5 | |
| 41 | 94.5 | |
| 42 | 94.5 | |
| 43 | 94.5 | |
| 44 | 94.5 | |
| 45 | 94.0 | |
| 46 | 94.0 | |
| 47 | 93.7 | |
| 48 | 93.7 | |
| 49 | 93.5 | |
| 50 | 93.0 |