Odysseys2
Agents · 2026-10-16
Carnegie Mellon University and Browser Use's benchmark of 100 long browser-agent tasks on the live web, constructed from the real request distribution of Browser Use users and graded against 8-12 rubric items each (982 in all) by a GPT-5.6 Luna judge that reads the action history, screenshots and delivered file contents. An attempt's final score is the share of rubric items met, set to zero when the judge flags suspected fabrication or reward hacking; the headline is the mean final score over one attempt per task.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 84.6 |
| GPT-5.6 Luna | 71.9 |
| Claude Opus 5.5 | 71.2 |
| Grok 4.7 | 68.2 |
| GPT-5.6 Sol | 62.0 |
| Claude Fable 5 | 61.5 |
| Claude Opus 5 | 58.3 |
| Grok 4.5 | 48.2 |
| GPT-5.6 Terra | 44.6 |
| Sonnet 5 | 42.3 |
| GPT-5.4 Mini | 24.1 |
| Gemini 3.1 Pro Preview | 10.8 |
| Gemini 3.8 Flash | 10.7 |
| Gemini 3.7 Flash | 5.8 |