Photo: Merlin Lightpainting / Pexels
The AI search company scored 80 on the new benchmark, a 47-point improvement over baseline performance and five points ahead of its nearest competitor.
Perplexity just claimed the top spot on a brand-new leaderboard designed to measure how well search APIs actually perform when plugged into AI agents. The Artificial Analysis Search Index, which launched on August 18, gave Perplexity’s medium Search API configuration the highest score of 80, putting it ahead of Parallel Search at 75 and Exa Search at 74.
To put that 80 in perspective: running the same tasks without any search tool at all produced a baseline score of just 33. That’s a 47-point lift, which is the kind of gap that makes you wonder why anyone would run an AI agent without search plugged in.
How the benchmark works #
The Artificial Analysis Search Index isn’t just one test. It combines three separate benchmarks: DeepSearchQA, BrowseComp, and AA-Omniscience, totaling 1,700 questions across varied research domains.
Every search API was evaluated against a fixed GPT-5.6 Luna model running within the Stirrup agent harness framework. Think of it as a controlled experiment where the model stays the same and only the search tool changes, isolating exactly how much value the search layer adds.
The index grades on three dimensions: answer quality, cost per task, and processing time.
Perplexity didn’t just win overall. Its variants, including medium, high, and low configurations, dominated the top rankings.
The economics of AI search #
Perplexity’s medium Search API runs $91.39 per 1,000 tasks when you add up the search cost ($62.30) and the model cost ($29.09). Average task completion clocks in at 28.0 seconds.
For context, the underlying Search API itself, which Perplexity launched on September 25, 2025, prices out at $5 per 1,000 requests for raw ranked web results. Those results pull from an index spanning hundreds of billions of pages. The total cost per task in the benchmark is higher because it includes the model inference costs needed to actually process and synthesize those search results into answers. During internal evaluations at launch, Perplexity recorded a median response time of 358 milliseconds.
What it means for the AI search landscape #
Artificial Analysis, founded in 2023, has built its reputation on creating standardized assessments for AI models and tools.
The competitive gap, while not enormous in absolute terms (five points separating first and second place), becomes more meaningful when paired with the cost data.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our