Browser Throughput Benchmark
Browser Throughput
Per-action throughput inside a running session — 50 sequential actions across 10 sessions per provider.
Want to see a provider added?
Methodology
What We Measure
Each session executes 50 sequential actions inside a single running browser: a repeating 10-step loop of navigate → waitForSelector → screenshot → textContent → click → waitForSelector → screenshot → textContent → goBack → waitForSelector, run 5 times against Wikipedia.
Every action is timed individually. A failed action does not abort the session — the loop continues and the failure is recorded. This lets us observe degradation patterns rather than discarding unlucky sessions.
Each provider runs 10 sessions per benchmark run. All sessions use stealth mode and a 1920×1080 viewport to match realistic agent conditions. Benchmarks run daily via GitHub Actions and results are committed to the public benchmarks repo.
How We Score
The Composite Score (0–100, higher is better) weights four axes that matter for agent workloads, then multiplies by success rate so flaky providers can't hide behind fast averages.
- • Actions/sec median: 40% — primary throughput signal
- • Task duration median: 25% — typical session cost
- • Task duration p95: 20% — tail consistency
- • Screenshot median: 15% — vision-agent proxy
The final score is multiplied by success rate — the fraction of sessions that completed all 50 actions. A provider that fails 1 in 10 actions is penalized proportionally.