Parallel clinches the top spot in the AI Search API, with Exa and Firecrawl following closely behind.

Perceive Beating AI News Flash: AI evaluation agency Artificial Analysis has introduced a Search Index, initiating scoring for the search API used by Agents. The search API serves as the interface for Agents to retrieve information online.

The evaluation consistently employs the GPT-5.6 Luna model and the same Harness framework, allowing a maximum of 25 Agent rounds per question, only swapping out the underlying search service. The initial test assessed 11 configurations from 7 manufacturers, with Parallel Advanced scoring 75 points, Exa Auto 74 points, and Firecrawl 73 points. A model that did not utilize any search received only 33 points.

The evaluation equally weighs DeepSearchQA, BrowseComp, and AA-Omniscience, testing deep search, multi-hop fact-finding, and knowledge-based questions respectively.

Surprisingly, Parallel Advanced incurred higher search costs compared to basic, yet the average token usage by the model decreased from around 33.9k to 16.9k. Consequently, the overall cost per question decreased from $0.11 to $0.084, with the score improving from 73 to 75. A slightly higher cost for the search API does not necessarily mean the Agent ends up being more expensive.

However, the score of 75 only represents search quality and does not account for the overall ranking based on quality, price, and speed. Artificial Analysis separately analyzes costs and latency, using a Pareto frontier to compare the trade-offs of different solutions.