AI agent benchmark compares 5 search APIs over MCP, with Parallel fastest and Keiro cheapest

Water_Law2005 · reddit · 2026-08-04

Benchmarked Keiro, Tavily, Exa, Parallel, and Valyu as web-search tools inside an AI agent over MCP.

Compared raw API speed, agent-call reliability, and answer quality scored on accuracy, completeness, and sourcing.
Results in the post's table:
Exa: 100% reliability, 2.4s agent speed, 1.3s raw speed, $7/1k, quality 8.5/10
Keiro: 100%, 5.5s, 1.9s, $0.25/1k, 8.5/10
Parallel: 100%, 1.4s, 2.1s, $3/1k, 8.5/10
Tavily: 82%, 2–3s, 2.5s, $8/1k, 7.5/10
Valyu: 40%, 6–12s, 4.8s, $1.50/1k, 7.5/10
A notable caveat: Exa and Parallel required exact tool names; natural language kept triggering the agent's built-in search.
Keiro had the broadest toolset (6 general-purpose tools), though two specialized tools had issues through this connector even though the direct API worked.

More from coding & agent