Jev + Mercury 2.5 + WebMCP aces browser benchmark at 1/112 the cost of GPT-6 Astra

The latest run on the open-source WindTunnel benchmark (nekuda-ai/WindTunnel) shows Jev paired with the low-cost model Mercury 2.5, operating the browser via the WebMCP protocol, completed 100% of 49 tasks at roughly 1/112 the model cost of GPT-6 Astra using computer use plus code execution, and substantially cheaper than screenshot-based computer use. The WebMCP benchmark team, which ran the test on its own benchmark, said the combo "basically blew up the benchmark."

Confirmed

• Jev + Mercury 2.5 + WebMCP achieved a 100% completion rate on 49 browser tasks, at about 1/112 the model cost of GPT-6 Astra's code-execution setup (m1/m5/m6).
• Division of labor: Jev can only pick from a discrete set of options (e.g., selecting the searchproducts tool) and cannot generate arbitrary text such as search terms; Mercury 2.5, which outputs 1000+ tokens per second, fills in tool parameters (m2/m3).
• Adding WebMCP raised the completion rate from 25/49 to full marks and cut costs by a further 18% (m3).
• WindTunnel compares different ways browser agents interact with websites; its core finding is that most tasks are bottlenecked on "picking the right next step" rather than generating tool parameters. WebMCP speeds browser agents up 2.5-7.5x and cuts costs 3-47x (m4).

Why it matters

The result challenges the assumption that browser agents need expensive large models doing computer use: since the hard part is decision-making rather than text generation, pairing a lightweight selection component with a cheap, low-latency LLM can slash cost and latency while maintaining or improving completion rates — a more economical, protocol-based path for agent-website interaction.

2026-09-18 ~ 2026-09-20 · 6 related posts

• [source] WebMCP Benchmark: Jev + Mercury 2.5 Solves 100% of Tasks at 245x Lower Cost Than Computer Use — hackgoofer · 2026-09-18
• WebMCP benchmark: Jev + Mercury 2.5 solves 100% of tasks at 112x lower cost than GPT-6 Astra — hardimanjames · 2026-09-19
• Jev + WebMCP solves 100% of benchmark tasks at 112x lower cost than GPT-6 Astra — QuanquanGu · 2026-09-19
• [source] Jev + Mercury 2.5 sweeps WebMCP benchmark 49/49 at ~112x lower cost than GPT-6 Astra — FinanceYF5 · 2026-09-20
• Jev picks tools but can't generate text—Mercury 2.5 fills parameters at 1000+ tokens/sec — FinanceYF5 · 2026-09-20
• [source] WindTunnel benchmark: WebMCP makes browser agents 2.5-7.5x faster, 3-47x cheaper — FinanceYF5 · 2026-09-20