BrowserGym AssistantBench: leaderboard
AssistantBench's realistic, time-consuming open-web tasks as run on ServiceNow's BrowserGym leaderboard. Scored as the benchmark's answer accuracy (%), computed on the tasks every listed agent finished; higher is better. Runs use AgentLab's GenericAgent scaffold unless the row names another framework.
Metric: Accuracy (%, GenericAgent unless named; tasks finished by every agent). Source: huggingface.co. Saturation forecast: Around September 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 Mini (2024-09-12) | 6.9 | #794 |
| 2 | Claude 3.5 Sonnet (20241022) | 5.2 | #635 |
| 3 | GPT-4o (2024-08-06) | 4.8 | #731 |
| 4 | Llama 3.1 405B | 3.9 | #1083 |
| 5 | Llama 3.1 70B | 2.8 | #1493 |
| 6 | GPT-4o Mini (2024-07-18) | 2.1 | #1512 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=browsergym-assistantbench · How It Works · Data refreshed daily, snapshot 2026-11-12.