Perplexity Search API Tops Agent Leaderboard

Perplexity has entered the search-for-agents market with its new Search API, which immediately claimed the top spot on the Artificial Analysis Search Index. Tested on the open-source Stirrup harness using GPT-5.6 Luna as the candidate model, all three context-size variants of the API outperformed previous leaders. The medium context variant scored 80, surpassing the previous high score of 75 held by Parallel advanced and Brave LLM context. The benchmark evaluates agent performance across DeepSearchQA, BrowseComp, and AA-Omniscience, with Perplexity showing its strongest lead in the multi-hop BrowseComp benchmark.

A key driver of this performance is Perplexity's flat pricing of $5.00 per 1,000 queries across low, medium, and high context settings. Because the API delivers compact payloads, it reduces the tokens the candidate model must process. This lowers model inference costs to between $0.028 and $0.034 per task, compared to the next-lowest provider at $0.036. When combining search API and model token costs, the total cost per task lands at approximately $0.091 for both medium and high variants, compared to Brave LLM context at $0.13 and Parallel advanced at $0.084.

The three variants reveal a clear scaling pattern. The low context variant scored 77, requiring 15.4 searches per task and costing about $0.105 because the agent ran extra searches to compensate for thin payloads. On BrowseComp, the low variant required 18.0 searches compared to 13.1 for the high variant. The high variant scored 79, requiring 11.4 searches per task. Latency also favored larger payloads, with medium and high taking 27 to 29 seconds per task, while the low variant took 36 seconds. Per-query speed was 1.1 seconds across all settings.

For AI practitioners, these results establish the medium context setting as the optimal default for agent architectures, as it delivers the highest quality score of 80 at the lowest total cost of $0.091. The high setting remains a viable option for complex, multi-hop reasoning tasks, while the low setting is inefficient and costly. By pushing the quality-cost Pareto frontier forward, Perplexity provides developers with a powerful tool to build faster, more economical search agents.