GAIA measures general-purpose agent performance on realistic multi-step tasks with tools, web navigation, and file operations. It is a practical benchmark for assistant workflows beyond single-turn QA.
Tasks require real tool use and sequential planning, limiting shortcut wins from pattern matching alone; three difficulty levels expose where agents fail as complexity rises.
Performance depends heavily on orchestration stack, tool configuration, and prompt policy; task count is smaller than mass-scale QA benchmarks, so confidence intervals are wider.
Higher is better. Accuracy percentage across all levels.
Max score: 100
Performance depends heavily on orchestration stack, tool configuration, and prompt policy.
| # | Model | Score |
|---|---|---|
| 1 | 89.1 | |
| 2 | 89.1 | |
| 3 | 88.1 | |
| 4 | 87.3 | |
| 5 | 87.3 | |
| 6 | 83.3 | |
| 7 | 82.0 | |
| 8 | 80.6 | |
| 9 | 79.4 | |
| 10 | 79.3 | |
| 11 | 77.8 | |
| 12 | 77.7 | |
| 13 | 77.6 | |
| 14 | 77.5 | |
| 15 | 77.3 | |
| 16 | 77.2 | |
| 17 | 75.8 | |
| 18 | 75.7 | |
| 19 | 74.8 | |
| 20 | 74.6 | |
| 21 | 74.5 | |
| 22 | 74.3 | |
| 23 | 74.3 | |
| 24 | 74.2 | |
| 25 | 74.2 | |
| 26 | 73.9 | |
| 27 | 73.4 | |
| 28 | 73.4 | |
| 29 | 72.6 | |
| 30 | 72.0 | |
| 31 | 71.8 | |
| 32 | 71.7 | |
| 33 | 71.1 | |
| 34 | 71.0 | |
| 35 | 70.9 | |
| 36 | 70.8 | |
| 37 | 70.5 | |
| 38 | 70.4 | |
| 39 | 70.4 | |
| 40 | 70.3 | |
| 41 | 70.2 | |
| 42 | 70.2 | |
| 43 | 69.7 | |
| 44 | 69.6 | |
| 45 | 69.4 | |
| 46 | 69.2 | |
| 47 | 68.9 | |
| 48 | 68.7 | |
| 49 | 68.4 | |
| 50 | 68.0 |