Artificial Analysis Puts 20 Search APIs to the Test, and Perplexity Comes Out on Top
AI agents that research, code, or answer questions need fresh data from the web. The layer that makes that possible is the search API, a service that returns web results to the model at query time. Artificial Analysis built a leaderboard that measures how well these APIs actually perform, ranking 20 products from 10 providers on answer quality, cost, and speed. The results give developers a concrete way to compare tools that differ in web index coverage, result formatting, and the tradeoffs they make between price and retrieval depth.
The core problem is that not all search APIs behave the same under the same agent. Some return rich snippets that let the model extract answers quickly. Others dump raw URLs and force the model to spend extra reasoning cycles parsing pages. The quality of the underlying web index varies. Pricing models range from flat-rate to per-query charges. Before this benchmark, developers had limited data to guide selection beyond vendor claims and informal testing.
Artificial Analysis tested each API by plugging it into the same agent setup. The candidate answer model is GPT-5.6 Luna (medium). The only variable is which search API the model calls. Three benchmarks measure answer quality: DeepSearchQA (900 tasks requiring multi-step research), BrowseComp (200 hard-to-find factual questions needing multi-hop browsing), and AA-Omniscience (600 factual questions balanced across six domains). Scores are combined into the Artificial Analysis Search Index on a 0 to 100 scale, giving a single comparable number across providers.
Perplexity Search Leads on Quality, But at a Price
Perplexity Search (medium) tops the leaderboard with an index score of 80, lifting quality by 47 points over the model-only baseline of 33. Perplexity Search (high) and Octen Search (highlights) follow at 79 and 77 respectively. The top three all score above 80 on DeepSearchQA, where broad research questions demand many searches. On BrowseComp, which tests obscure multi-hop facts, Perplexity Search (medium) hits 87 accuracy. Firecrawl Search scores highest on AA-Omniscience at 73, showing strength on pure factual recall.
The model-only baseline (GPT-5.6 Luna with no search) scores 33 on the index, demonstrating that search APIs provide substantial value for knowledge-intensive tasks. The gap between the best and worst search-enabled providers is 9 points (80 vs. 71), but the gap between any search provider and no search is 38 to 47 points.
TinyFish Offers Free Search, While Octen Wins on Speed
Cost varies widely. TinyFish Search (web) charges nothing for search per 1,000 tasks, though the model token cost still applies at $34.55 per 1,000 tasks. Octen Search (highlights) is the cheapest paid option at $9.07 per 1,000 tasks for search. Perplexity Search (low) costs $77.20 per 1,000 tasks, the most expensive search component. Model token costs range from $27.58 (Perplexity low) to $69.94 (You.com highlights), because different APIs return different payload sizes and trigger different amounts of reasoning.
On latency, Octen Search (highlights) is the fastest at 17.2 seconds per task, and also the fastest per individual search query at 0.21 seconds. Perplexity Search (low) is the slowest at 37.4 seconds per task. Firecrawl Search takes 63.2 seconds per task, though it scores well on factual accuracy. The correlation between speed and quality is not direct. A provider can be fast per call but still add total time if the model makes more searches against it.
What the Benchmarks Actually Measure
DeepSearchQA asks broad research questions where answers are lists of items. An LLM grader scores each answer using F1 over the answer items. BrowseComp uses hard, multi-hop questions with exact match grading. AA-Omniscience tests factual recall across six domains. Each benchmark targets a different aspect of search-augmented generation. The index averages all three, so providers that perform well across the board rank higher.
The task examples illustrate the difficulty. One BrowseComp question asks for a structure in Eastern Australia rebuilt in 2016, longer than 50 meters, visible from a similar structure, and hosting a yearly dinner. One AA-Omniscience question asks about a jazz band's origin in a Karen Russell short story. These are not simple keyword lookups. They require search APIs to return results that let the model reason across multiple hops of information.
How to Pick the Right Search API
The choice depends on what the agent prioritizes. For maximum answer quality, Perplexity Search (medium) is the clear leader. For cost-sensitive applications, TinyFish Search (web) or Octen Search (highlights) offer the lowest search fees. For real-time use cases where latency matters, Octen Search (highlights) and You.com Search (highlights) are the fastest. The scatter plots in the benchmark data show the Pareto frontier for quality-cost and quality-latency tradeoffs, helping developers find providers that sit on the efficient edge.
The benchmark also reveals that search quality affects downstream costs. Lower-quality search results lead to more retries and higher reasoning token usage, inflating the total model cost. Providers that return well-structured, relevant results reduce the number of searches the model needs to run, which saves both time and money.
A Moving Target Worth Tracking
Search APIs improve their indexes, pricing, and result formats regularly. The benchmark data is current as of September 8, 2026, and the leaderboard will update as providers change. For teams building AI agents that need web access, this data provides a starting point for evaluation rather than relying on vendor demos. The full methodology details scoring, latency measurement, and cost breakdowns on the Artificial Analysis site.