ARC-AGI-2 is the harder successor to ARC-AGI-1 for abstract puzzle generalization. It is aimed at separating frontier systems after v1 saturation trends.
Harder task distribution preserves separation among frontier systems; semi-private setup continues to limit straightforward benchmark gaming.
Still measures a narrow abstraction domain rather than end-to-end product utility; reported outcomes can shift with retry budgets and evaluation harness choices.
Higher is better. Percentage of tasks solved correctly.
Max score: 100
Still measures a narrow abstraction domain rather than end-to-end product utility.
| # | Model | Score |
|---|---|---|
| 1 | 100.0 | |
| 2 | 100.0 | |
| 3 | 99.9 | |
| 4 | 99.7 | |
| 5 | 99.5 | |
| 6 | 98.8 | |
| 7 | 97.1 | |
| 8 | 95.8 | |
| 9 | 94.6 | |
| 10 | 94.2 | |
| 11 | 94.2 | |
| 12 | 93.3 | |
| 13 | 93.3 | |
| 14 | 92.9 | |
| 15 | 92.7 | |
| 16 | 92.5 | |
| 17 | 92.5 | |
| 18 | 92.1 | |
| 19 | 92.1 | |
| 20 | 92.1 | |
| 21 | 91.7 | |
| 22 | 91.7 | |
| 23 | 91.7 | |
| 24 | 90.4 | |
| 25 | 90.0 | |
| 26 | 89.4 | |
| 27 | Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 5 Fallback) source arc_agi Evidence observed Nov 22, 2026 | 89.2 |
| 28 | 89.2 | |
| 29 | 88.3 | |
| 30 | 87.5 | |
| 31 | 86.7 | |
| 32 | 86.7 | |
| 33 | 85.4 | |
| 34 | 85.4 | |
| 35 | 85.0 | |
| 36 | 84.6 | |
| 37 | 84.6 | |
| 38 | 84.4 | |
| 39 | 84.2 | |
| 40 | 83.3 | |
| 41 | 83.3 | |
| 42 | 82.9 | |
| 43 | 82.5 | |
| 44 | 82.1 | |
| 45 | 79.3 | |
| 46 | 79.2 | |
| 47 | 78.3 | |
| 48 | 78.1 | |
| 49 | 78.0 | |
| 50 | 77.5 |