Jev + Mercury 2.5 sweeps WebMCP benchmark 49/49 at ~112x lower cost than GPT-6 Astra

FinanceYF5 · x · 2026-09-20

A breakdown of an agent architecture built on interaction method rather than raw model power:

Jev can only pick tools from discrete options; it can't generate arbitrary text like search queries.
Mercury 2.5 fills that gap by generating tool parameters at 1000+ tokens/sec, fast and cheap.
WebMCP turns website functionality into clean, directly callable tools.

Results: Jev alone via browser DOM completed only 25/49 tasks; with WebMCP it hit 49/49 — nearly double — while model cost dropped another 18%. Per the thread, the combo costs 112x less than GPT-6 Astra with code execution and 245x less than Astra driving a browser via screenshots. The takeaway: how an agent interacts with websites matters more than the model itself.

Original post →

Jev + Mercury 2.5 sweeps WebMCP benchmark 49/49 at ~112x lower cost than GPT-6 Astra

More from coding & agent