# Perplexity takes all three top spots in Artificial Analysis' search API test > Source: > Published: 2026-08-28 20:19:58+00:00 # Perplexity takes all three top spots in Artificial Analysis' search API test **Its medium setting scored 80 at about $0.091 per task, beating rivals while landing in the middle of the latency table.** By [Ryan Merket](/author/ryan-merket) ยท Published Primary source: [Artificial Analysis on X](https://x.com/ArtificialAnlys/status/2093427938968666138) ## Why it matters AI agents increasingly depend on search providers as infrastructure. Perplexity's sweep shows that retrieval quality can change agent accuracy sharply, while medium context can outperform larger payloads at the same request price. Perplexity took first, second and third place in the [Artificial Analysis Search Index](https://artificialanalysis.ai/agents/search-api), giving its Search API the strongest quality scores in a controlled test of the retrieval systems feeding web data to AI agents. The August 27th leaderboard result puts a third-party score behind a core bet made by [Aravind Srinivas (@AravSrinivas)](https://x.com/AravSrinivas), the former OpenAI researcher who co-founded Perplexity in 2022 with Denis Yarats, Johnny Ho and Andy Konwinski. Perplexity has built its products around retrieving current web information and handing that material to language models, first through its consumer answer engine and later through APIs sold to developers. [Artificial Analysis reported the results](https://x.com/ArtificialAnlys/status/2093427938968666138) on August 28th after testing one Perplexity Search API endpoint at its low, medium and high context settings. Those settings control how much content is extracted from each result page and returned to an application. The medium setting led the index with a score of 80, followed by high at 79 and low at 77. The next-best results, Parallel Search's advanced setting and Brave Search's LLM context mode, each scored 75. You.com Search's highlights mode and Exa Search's auto mode scored 74. More context did not produce the best result. Perplexity's medium setting beat high by one point and low by three, suggesting that selecting the right amount of page content matters as much as returning the largest possible payload. Excess material can raise downstream model usage or make the relevant passage harder for an agent to isolate. That finding tracks with Perplexity's own work on extraction. In a March 11th product update, Perplexity said it had been optimizing Search API snippets for relevance and size, including systems that label which spans of a document directly answer a query. The [current API documentation](https://docs.perplexity.ai/api-reference/search-post) describes low as short, query-relevant passages, medium as a balanced amount of content and high as detailed content. ### Quality came without a speed win Artificial Analysis calculated an average total cost of roughly $0.091 per task for both the medium and high settings. Low cost about $0.105 per task. Those totals combine Search API charges with the inference cost of the model operating the agent. Perplexity charges [$5 per 1,000 Search API requests](https://docs.perplexity.ai/docs/getting-started/pricing), regardless of the selected context setting. Richer context therefore carries no higher per-request search fee. It can still change the final bill by affecting how many searches the agent performs and how many tokens the answer model processes. Perplexity's medium result extended the benchmark's quality-cost frontier from a previous maximum score of 75 to 80 at about $0.091 per task. Parallel Search's fast setting remained cheaper at about $0.068 per task, although it scored 73. Latency produced a less decisive result. Perplexity's medium and high settings took 28.3 and 29.1 seconds per task, respectively. Low took 36.6 seconds. Parallel Search's fast setting completed tasks in 19 seconds, Exa Search's instant mode took 19.7 seconds and You.com's highlights mode took 20.2 seconds. The timing metric includes both search calls and model processing. A provider can return an individual query quickly and still produce a slower task if the agent needs additional searches or spends longer reasoning over the retrieved material. ### A controlled test of the retrieval layer Artificial Analysis designed the [methodology](https://artificialanalysis.ai/methodology/search-api) to hold the rest of the agent constant. Each provider was paired with [GPT-5.6 Luna](/models/openai/gpt-5.6-luna:batch) at medium reasoning effort inside Artificial Analysis' open-source Stirrup harness. The agent received web search and page-fetching tools, a maximum of 25 turns and up to 10 results per search. The index equally weights three evaluations: 900 DeepSearchQA tasks, 600 held-out factual questions from AA-Omniscience and a difficult 200-question subset of BrowseComp. The same model also graded responses under benchmark-specific rules. That design isolates the search provider more cleanly than tests that compare complete products running different models and agent instructions. It still measures one model, one harness and one collection of tasks. Production results can shift with another model, a different context strategy or workloads concentrated in areas such as coding, commerce or local search. Perplexity nevertheless swept the quality table under identical conditions, and the medium setting delivered the best balance. For developers, the useful result is narrower than a blanket claim of search superiority: Perplexity supplied the most effective retrieval payload to this particular agent, and sending the agent the maximum available context was unnecessary.