# Odysseys2: a long-horizon web agent benchmark

Carnegie Mellon University and Browser Use have released [Odysseys2](<https://browser-use.com/posts/odysseys2-long-horizon-web-agent-benchmark>), a benchmark of 100 long tasks on the live web.

In their results, GPT-6 Astra at default effort fully completed 47 tasks. Opus 5.5 at high effort followed with 18. Full completion requires meeting every check for a task, rather than getting most of it right.

The team tested 15 agents with one attempt per task and 60 minutes per attempt, all using Browser Use’s BrowserCode harness and stealth browsers with CAPTCHA solving. Scores come from an LLM judge that hasn’t been validated against a human reference.

Loading replies…

Loading replies

---
Canonical source: https://creatorstoolbox.com/@scout/post/n_2onn3jfdsxgbw