# Owl Browser Benchmarks ## Agent Efficiency Benchmark Tokens per task across MCP browser servers. Six browsing tasks run 60 times per stack (30 via Claude Code CLI, 30 via Codex CLI) against the same pinned local model (qwen3.6-27b at Q4_K_M). Every stack answers every task correctly, so token cost and call count are what separate them. | Metric | Owl Browser MCP | Chrome DevTools MCP | Playwright MCP | |---|---|---|---| | Median tokens per task | 38,590 | 71,437 | 73,420 | | Median tool calls per task | 2.0 | 3.0 | 4.0 | | Correctness | 100% | 100% | 100% | Data source: https://owlbrowser.net/detection-reports/agentic.json ## Built-in NLA Benchmark Plain-English tasks with no agent CLI. The same six tasks given to `browser_nla`, the agentic observe-act loop that runs inside Owl Browser. | Metric | Owl Browser Built-in NLA | |---|---| | Median tokens per task | 4,844 | | Median steps per task | 2 | | Tasks answered correctly | 18/18 | Note: These token counts are not comparable with the MCP figures because there is no external agent CLI re-sending a tool catalogue and the whole conversation every turn. Data source: https://owlbrowser.net/detection-reports/nla.json ## Anti-bot Detection Benchmark Automated fingerprint comparison of Owl Browser against default Playwright and Puppeteer on CreepJS, BrowserScan, and reCAPTCHA v3. Owl Browser applies fingerprint spoofing at the Chromium C++ source level rather than through JavaScript injection, so spoofed values come from the same Blink code paths as genuine ones. Data source: https://owlbrowser.net/detection-reports/report.json ### Fingerprint Uniqueness Playwright and Puppeteer expose identical device fingerprints, leaking the real machine. Owl Browser generates unique fingerprints per profile at the C++ source level.