Best web search API by F1 in this benchmark: Parallel basic at 46.5%. Fastest search: SERP (RapidAPI), 2.93s (0.4% F1). Lowest recorded search price: TinyFish, $0.00 per 1,000 queries within its free-tier limits.
Choose the search engine behind an AI agent that plans several searches and combines the evidence. Compare 12 providers across 20 configurations, including Exa, Tavily, Brave, Parallel and Firecrawl, on the same questions.
Updated 20 September 2026. Last measured 14 September 2026.
Mean F1 on 45 company-discovery questions with multiple constraints, using the same agent and search results. F1 balances precision and recall; it is distinct from percentage accuracy on a single-answer test. Prices are the benchmark-recorded rate basis, with the same configurations used for the quality results.
| Rank | Provider & configuration | F1 | Mean search time | Recorded $ / 1,000 queries | Median task time | Median task cost | Official docs |
|---|---|---|---|---|---|---|---|
| 1 | Parallel basicparallel-basicPOST /v1/search mode=basic | 46.5% | 19.77s | $5.00mode=basic · 10 results | 67.63s | $1.12 | Official docs ↗ |
| 2 | Exa deepexa-deepPOST /search type=deep | 45.4% | 44.68s | $12.00type=deep | 89.50s | $0.72 | Official docs ↗ |
| 3 | Parallel advancedparallel-advancedPOST /v1/search mode=advanced | 44.2% | 36.58s | $5.00mode=advanced · 10 results | 83.43s | $0.62 | Official docs ↗ |
| 4 | Exa instantexa-instantPOST /search type=instant | 43.3% | 6.21s | $7.00type=instant · up to 10 results | 49.83s | $0.65 | Official docs ↗ |
| 5 | Linkup fastlinkup-fastPOST /v1/search depth=fast | 41.1% | 17.52s | $5.00depth=fast · searchResults | 64.20s | $0.92 | Official docs ↗ |
| 6 | Tavily advancedtavily-advancedPOST /search search_depth=advanced | 41.1% | 46.58s | $16.002 credits · $0.008 PAYG | 92.17s | $1.03 | Official docs ↗ |
| 7 | Linkup standardlinkup-standardPOST /v1/search depth=standard | 40.6% | 25.10s | $5.00depth=standard · searchResults | 72.29s | $0.94 | Official docs ↗ |
| 8 | Youextraction_mode=highlightsPOST /v1/search | 38.6% | 7.72s | $5.00highlights included in Web Search | 48.16s | $0.97 | Official docs ↗ |
| 9 | Youextraction_mode=highlights · knowledge=corePOST /v1/search | 38.1% | 8.88s | $5.00highlights · knowledge=core · same Web Search price assumption | 47.64s | $0.90 | Official docs ↗ |
| 10 | Parallel fastparallel-fastPOST /v1/search mode=fast | 38.0% | 12.17s | $1.00mode=fast · 10 results | 53.93s | $0.46 | Official docs ↗ |
| 11 | Perplexitysearch_context_size=lowPOST /search | 37.8% | 14.11s | $5.00Search API · POST /search | 48.94s | $0.33 | Official docs ↗ |
| 12 | Tavily basictavily-basicPOST /search search_depth=basic | 36.5% | 20.09s | $8.001 credit · $0.008 PAYG | 63.04s | $0.61 | Official docs ↗ |
| 13 | Parallel turboparallel-turboPOST /v1/search mode=turbo | 34.7% | 5.08s | $1.00mode=turbo · 10 results | 46.40s | $0.42 | Official docs ↗ |
| 14 | Nimblesearch_depth=standard · full_content=false · focus=generalPOST /v2/search | 30.7% | 12.22s | $5.00search_depth=standard · full_content=false | 52.97s | $0.47 | Official docs ↗ |
| 15 | FirecrawlfirecrawlPOST /v2/search | 30.4% | 39.44s | $5.002 credits / 10 results | 74.98s | $0.28 | Official docs ↗ |
| 16 | Brave SearchbraveGET /res/v1/web/search | 28.0% | 7.60s | $5.00Recorded search-request rate; see official terms. | 43.52s | $0.27 | Official docs ↗ |
| 17 | TinyFishtinyfishGET api.search.tinyfish.ai | 26.6% | 27.34s | $0.0030 req/min cap | 64.24s | $0.21 | Official docs ↗ |
| 18 | Nimblesearch_depth=lite · full_content=false · focus=generalPOST /v2/search | 24.1% | 34.43s | $1.10search_depth=lite · full_content=false | 67.17s | $0.22 | Official docs ↗ |
| 19 | Seltzseltz-companiesPOST /v1/search scope=companies | 14.5% | 8.98s | $5.00Recorded search-request rate; see official terms. | 55.16s | $1.75 | Official docs ↗ |
| 20 | SERP (RapidAPI)serpGET google-search74.p.rapidapi.com | 0.4% | 2.93s | $3.00Pro overage $0.003/request | 31.38s | $0.10 | Official docs ↗ |
Here, an AI search engine means a programmatic web search service that feeds an agent. Compare the API's returned evidence and the agent's resulting answer. Keep consumer chat products and model-generated answer services in their own evaluation, because those products combine retrieval with a different answering system.
An agent may need to discover candidates, refine a query and check several constraints before answering. This board tests 45 company-discovery questions with the same agent and tool budget for each provider. F1 balances precision and recall across the returned company sets. A broader result set helps only when it contains the correct matches.
A quick search call can still lead to a long reasoning loop. The table therefore shows both mean search time and median task time, together with F1. Read the fastest result alongside its quality score. The agent uses returned search evidence in this board; a separate benchmark adds page scraping to measure that workflow.
Use each benchmark to answer the question it measured. Openbenchmarks keeps factual-search accuracy, developer task completion and multi-search F1 separate. AIMultiple evaluates returned-result quality and relevance. Artificial Analysis combines several agent-search evaluations into its Search Index. Compare their methods and source results before treating a rank in one study as a rank in another.
Mean F1 on 45 company-discovery questions with multiple constraints, using the same agent and search results. F1 balances precision and recall; it is distinct from percentage accuracy on a single-answer test. Each row is one measured configuration. Missing measurements appear as a dash.
Open benchmark code ↗ · Public dataset sample ↗ · All benchmark methods →
Compare multi-search and search-plus-scrape results →
Other studies: AIMultiple's returned-result evaluation ↗ · Artificial Analysis Search Index methodology ↗. Their datasets and scoring differ from the F1 board above.
Best web search API by F1 in this benchmark: Parallel basic at 46.5%. Fastest search: SERP (RapidAPI), 2.93s (0.4% F1). Lowest recorded search price: TinyFish, $0.00 per 1,000 queries within its free-tier limits.
Mean F1 on 45 company-discovery questions with multiple constraints, using the same agent and search results. F1 balances precision and recall; it is distinct from percentage accuracy on a single-answer test.
The recorded search endpoint's request price. Model-token spend, optional scraping and plan conditions are separate. For rows with task measurements, median task cost is shown in its own column. Check the linked provider terms for current pricing.