WindTunnel benchmark: WebMCP makes browser agents 2.5-7.5x faster, 3-47x cheaper
FinanceYF5 · x · 2026-09-20
WindTunnel, an open benchmark from nekuda-ai, compares how browser agents interact with websites. Key insight: the hard part is choosing the right next step, not generating tool arguments—WebMCP compresses clicks, forms and menus into one explicit tool call, shrinking the decision space. On the leaderboard Jev + Mercury 2.5 (WebMCP) tops at 96.5 (49/49 tasks), with GPT-5.6 Luna native WebMCP at 91.5. Versus the median screen-driving agent (44/49), WebMCP setups score 27-50% higher, finish 2.5-7.5x faster, and cost 3-47x less per task. Tests and methods are open-sourced and reproducible.
More from coding & agent
- MCP memory server with per-app revokable grants instead of shared tokens — EntrepreneurOk5764 · 2026-11-04
- 'Evals are fuzzy unit tests': Tejas Kumar's AI Engineer talk on evals gets peer recommendation — TejasKumar_ · 2026-11-04
- SOMA hits 1,000 downloads: Claude Code designs a bookshelf in CAD, self-checks its work — Lazar988 · 2026-11-04
- Document extraction past the POC: layout drift breaks pipelines, and parser model swaps add reprocessing pain — Mission-Leave-6551 · 2026-11-04
- SpecJudge author argues LLM-as-judge tools should never ask models for self-reported confidence — jokiruiz · 2026-11-04
- SOMA, the Claude Code CAD app, passes 1,000 users with a 9-minute bookshelf design walkthrough — Pr8tham · 2026-11-04