New benchmark ranks search APIs for AI agents on quality, cost, and speed

Details

Artificial Analysis has published the Search Index, a head-to-head benchmark for search APIs used in AI agent pipelines. Seven providers were evaluated — Parallel, Exa, Firecrawl, Bing, You.com, Tavily, and Brave — across three dimensions: result quality, cost per query, and latency. Parallel topped the composite with a score of 75, followed by Exa at 74 and Firecrawl at 73. The evaluation used GPT-5.6 Luna as the judgment model and combined three benchmarks: DeepSearchQA for knowledge retrieval, BrowseComp for complex research tasks, and AA-Omniscience for broad information coverage. A key finding is that higher-quality search has a counterintuitive cost relationship: Parallel's better result summarization reduces downstream LLM token consumption by more than 40%, making its /bin/bash.084 all-in cost competitive with cheaper providers that require more tokens to compensate for lower-quality results. Exa's /bin/bash.11 total cost sits above Parallel's despite competitive quality scores, suggesting search API selection should factor total pipeline cost, not just per-query price. The benchmark is the first systematic evaluation tool for this component of the agentic stack, filling a gap practitioners have been working around with ad hoc testing.