WebArena measures autonomous web-task completion across realistic multi-site workflows. It is a practical signal for browser-agent navigation, form filling, and stateful planning.
Execution-based tasks on realistic interfaces test planning and recovery from UI errors; multi-domain setup reduces overfitting to one site pattern.
Results are highly sensitive to browser tooling and agent policy, not just core model quality; self-hosted infrastructure adds setup friction and reproducibility variance.
Higher is better. Task success rate percentage.
Max score: 100
Results are highly sensitive to browser tooling and agent policy, not just core model quality.
| # | Model | Score |
|---|---|---|
| 1 | 74.3 | |
| 2 | 64.5 | |
| 3 | 64.0 | |
| 4 | 61.1 | |
| 5 | 49.2 | |
| 6 | 38.3 | |
| 7 | 37.3 | |
| 8 | 36.2 | |
| 9 | 33.5 | |
| 10 | 32.4 | |
| 11 | 32.4 | |
| 12 | 31.6 | |
| 13 | 30.3 | |
| 14 | 29.4 | |
| 15 | 26.1 | |
| 16 | 26.1 | |
| 17 | 26.1 | |
| 18 | 26.1 | |
| 19 | 25.2 | |
| 20 | 24.3 | |
| 21 | 24.0 | |
| 22 | 23.5 | |
| 23 | 23.4 | |
| 24 | 23.2 | |
| 25 | 22.4 | |
| 26 | 22.4 | |
| 27 | 21.0 | |
| 28 | 20.9 | |
| 29 | 20.1 | |
| 30 | 19.4 | |
| 31 | 19.3 | |
| 32 | 19.3 | |
| 33 | 19.0 | |
| 34 | 18.8 | |
| 35 | 18.7 | |
| 36 | 18.6 | |
| 37 | 18.6 | |
| 38 | 18.5 | |
| 39 | 18.5 | |
| 40 | Gemini 2.5 Flash-Lite Preview (Sep '25) (Non-reasoning) source benchpress Evidence observed Apr 13, 2026 | 18.5 |
| 41 | 18.2 | |
| 42 | 18.1 | |
| 43 | 18.0 | |
| 44 | 17.9 | |
| 45 | 17.3 | |
| 46 | 17.2 | |
| 47 | 16.9 | |
| 48 | 16.9 | |
| 49 | 16.8 | |
| 50 | 16.7 |