Internal Bench Hard
Browser Use: 82% of tasks solved at 17¢ each. That is 20 points better than Opus 5, which costs 20× more per solved task.
Toolstrict accuracyCost
What it measures
Our hardest internal set: 106 tasks that a careful human can finish but most agents cannot. Every model runs on the same harness, and we count a task solved only on a strict match, so partial credit earns nothing.
How it was run
- Tasks: 106 hard tasks on live websites, none removed.
- Cost: total recorded spend for the run divided by the number of tasks actually solved.
- Scoring: strict — a task counts only when the final answer or end state is exactly right.