Best web search API by accuracy (task completion) in this benchmark: Perplexity at 77.3%. Fastest search: Parallel turbo, 333ms (64.7% task completion). Lowest recorded search price: TinyFish, $0.00 per 1,000 queries within its free-tier limits.
Choose the search tool an agent can call to find a useful source and finish a task. Compare 10 providers across 14 configurations, including Exa, Tavily, Brave, Parallel and Firecrawl, on the same questions.
Updated 20 September 2026. Last measured 15 September 2026.
Task completion on 100 developer questions with the same agent using search results. Each task requires a passing patch and a citation from the run's sources. Search latency and total task latency are shown separately. Prices are the benchmark-recorded rate basis, with the same configurations used for the quality results.
| Rank | Provider & configuration | Task completion | Mean search time | Recorded $ / 1,000 queries | Median task time | Median task cost | Official docs |
|---|---|---|---|---|---|---|---|
| 1 | Perplexitysearch_context_size=low/search | 77.3% | 957ms | $5.00Search API · POST /search | 18.31s | $0.06 | Official docs ↗ |
| 2 | Firecrawlfirecrawl-search/v2/search | 70.3% | 2.87s | $5.002 credits / 10 results | 27.95s | $0.06 | Official docs ↗ |
| 3 | Parallel fastmode=fast/v1/search | 66.7% | 953ms | $1.00mode=fast · 10 results | 22.09s | $0.05 | Official docs ↗ |
| 4 | Exa fasttype=fast/search | 66.3% | 626ms | $7.00type=fast · up to 10 results | 19.97s | $0.10 | Official docs ↗ |
| 5 | Parallel turbomode=turbo/v1/search | 64.7% | 333ms | $1.00mode=turbo · 10 results | 18.69s | $0.06 | Official docs ↗ |
| 6 | Exa instanttype=instant/search | 61.3% | 447ms | $7.00type=instant · up to 10 results | 20.58s | $0.10 | Official docs ↗ |
| 7 | TinyFishfree · 30 req/min capapi.search.tinyfish.ai | 59.3% | 2.15s | $0.0030 req/min cap | 27.84s | $0.03 | Official docs ↗ |
| 8 | Tavily basicsearch_depth=basic/search | 51.0% | 1.72s | $8.001 credit · $0.008 PAYG | 26.40s | $0.10 | Official docs ↗ |
| 9 | Nimblesearch_depth=lite · full_content=false · focus=general/v2/search | 46.7% | 3.15s | $1.10search_depth=lite · full_content=false | 32.72s | $0.04 | Official docs ↗ |
| 10 | Linkup fastdepth=fast · outputType=searchResults/v1/search | 43.3% | 1.39s | $5.00depth=fast · searchResults | 26.76s | $0.10 | Official docs ↗ |
| 11 | Nimblesearch_depth=standard · full_content=false · focus=general/v2/search | 42.3% | 878ms | $5.00search_depth=standard · full_content=false | 25.27s | $0.08 | Official docs ↗ |
| 12 | Youextraction_mode=highlights/v1/search | 41.3% | 596ms | $5.00highlights included in Web Search | 23.93s | $0.10 | Official docs ↗ |
| 13 | Youextraction_mode=highlights · knowledge=core/v1/search | 38.3% | 677ms | $5.00highlights · knowledge=core · same Web Search price assumption | 21.05s | $0.10 | Official docs ↗ |
| 14 | Bravecount=10 · result_filter=web/res/v1/web/search | 38.0% | 523ms | $5.00Recorded search-request rate; see official terms. | 19.87s | $0.08 | Official docs ↗ |
An agent tool connects a question to sources the model can use. Start with whether those sources let the agent finish the job, then compare the time and cost of each search call. In this benchmark, every provider gets the same agent and developer questions. Accuracy means task completion: the resulting patch must pass its checks and cite a source returned during the run.
Choose the output contract before choosing a vendor. Ranked links and excerpts fit an agent that reasons over sources itself. Scraped page text adds context when a result excerpt is insufficient. A generated answer adds a separate model and its own scoring problem. The table here compares agents using search results; the linked search-plus-scrape board measures the expanded workflow separately.
Use the official endpoint and configuration in the table. Return source URLs alongside text, retain those URLs through the agent loop, and record the number of searches and tokens used to complete each task. Compare providers on the same questions, result limit and model. Check citation support and completion quality before choosing the fastest row.
A task can make several searches. Search spend equals search calls multiplied by the per-call rate; add model tokens and any scrape charges to estimate the full task bill. The table keeps the recorded price per 1,000 queries separate from measured median task cost. Free-tier limits and paid-plan conditions appear beside the price.
Task completion on 100 developer questions with the same agent using search results. Each task requires a passing patch and a citation from the run's sources. Search latency and total task latency are shown separately. Each row is one measured configuration. Missing measurements appear as a dash.
Open benchmark code ↗ · Public dataset sample ↗ · All benchmark methods →
Compare the developer search and search-plus-scrape boards →
Best web search API by accuracy (task completion) in this benchmark: Perplexity at 77.3%. Fastest search: Parallel turbo, 333ms (64.7% task completion). Lowest recorded search price: TinyFish, $0.00 per 1,000 queries within its free-tier limits.
Task completion on 100 developer questions with the same agent using search results. Each task requires a passing patch and a citation from the run's sources. Search latency and total task latency are shown separately.
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.