Browser Agent Evals

readme.md →

How today’s models perform on computer use benchmarks, compared on accuracy, cost, and speed.

10 runs, ranked by accuracy

accuracy per model

accuracy
project=stagehand-devbenchmark=onlineMind2Webrows=10/10providers=3harnesses=2updated=2026-11-18T22:39:35Zrun=stagehand-evals

10 runs

accuracy vs cost per task

accuracy
claude-opus-4-8hybrid1st84.1%260s$1.156
gpt-5.5hybrid2nd72.3%285s$1.730
gemini-2.5-computer-use-preview-10-2025cua3rd71.3%180s3rd$0.273
claude-opus-4-7cua66.3%3rd142s$1.221
gemini-3-flash-previewcua65.3%209s2nd$0.0626
claude-sonnet-4-6cua62.7%387s$1.850
claude-sonnet-4-5cua57.3%339s$1.498
claude-haiku-4-5cua54.0%213s$0.468
gpt-5.4cua40.7%2nd108s$0.474
gpt-5.4-minicua24.3%1st62s1st$0.0578

*these results were collected on stagehand v3 and are kept for reference. for current numbers see

Run your own evals

same harness, your models, your benchmark

readme.md →

Want to get your model or harness evaluated?

talk to an engineer →

Your agent should be able to use a browser like you do.

The SDK for browser agents