Best by accuracy
- Accuracy
- 77.3%
- Mean search time
- 957ms
- Search price / 1,000 queries
- $5.00
Recorded search API price. Model and scrape costs are separate. Official docs ↗ · Pricing ↗
Independent comparison · 2026
Best Web Search APIs for RAG (2026): Search Retrieval Accuracy, Speed & Cost
Best web search API for RAG by accuracy: Perplexity, 77.3% at $5.00/1,000 queries. Fastest: Parallel turbo, 333ms. Cheapest: TinyFish, $0.00/1,000 queries within free-tier limits.
Compare the best web search APIs from 10 providers, including Exa, Tavily, Brave, Parallel and Firecrawl, for retrieval-augmented generation (RAG) and AI apps. We test 14 configurations with the same agent on 100 questions.
Fastest search
- Accuracy
- 64.7%
- Mean search time
- 333ms
- Search price / 1,000 queries
- $1.00
Recorded search API price. Model and scrape costs are separate. Official docs ↗ · Pricing ↗
Cheapest search
- Accuracy
- 59.3%
- Mean search time
- 2.15s
- Search price / 1,000 queries
- $0.00
Free-tier limits apply. Model and scrape costs are separate. Official docs ↗ · Pricing ↗
Compare all 14 search configurations → · Open benchmark code ↗ · Public dataset sample ↗
Updated 26 September 2026. Last measured 15 September 2026.
Web search for RAG: search retrieval accuracy, speed and cost
Search accuracy 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 ↗ |
From search result to a grounded RAG answer
| Stage | Keep or measure | How to use this comparison |
|---|---|---|
| Retrieve | Query, endpoint mode, result URLs, passages and search latency | Compare the same search configurations and request times in the main table. |
| Select context | Evidence coverage, source diversity and tokens passed to the model | Test your chunking and reranking on the same retrieved results. |
| Generate and cite | Answer correctness and support for each cited claim | Developer task completion measures successful evidence use in the tested agent. Evaluate your own answer format separately. |
| Operate | Search calls per answer, model spend and end-to-end time | Keep per-query search pricing separate from the full application bill. |
For direct factual retrieval, also see the search accuracy comparison. For an agent that assembles evidence across many searches, see the multi-search comparison.
Choose the best web search API for your application
What this retrieval benchmark tells a RAG builder
The main table measures a developer agent that retrieves documentation and uses that evidence to solve a task. Accuracy here means grounded task completion: a patch must pass its checks and cite a source from the run. That is useful evidence for technical RAG and agentic retrieval. General chatbot answer quality still needs evaluation on your own questions and generation pipeline.
Keep the source attached to every retrieved passage
Retain the source URL, title, passage and retrieval time together. Feed evidence into the answer prompt with stable citation identifiers, then check whether each final claim is supported by its cited passage. A high search score is one part of this pipeline; chunk selection, context limits and the answer model also influence the result.
Compare retrieval latency with answer latency
Mean search time measures the API call. Median task time includes repeated searches and the agent’s work. For a RAG assistant, measure retrieval, reranking and generation separately, then track end-to-end time. Choose a configuration that meets the accuracy level your application needs before trading quality for a shorter response.
Estimate the cost of a grounded answer
Search cost per answer equals the number of search calls multiplied by the per-call price. Add input and output tokens, reranking and any scraping. The table separates recorded search price per 1,000 queries from median full-task cost, so a cheap API request and a cheap completed answer remain distinct.
Test the RAG pipeline on your own questions
Build a held-out set covering fresh facts, ambiguous questions, missing evidence and conflicting sources. Inspect retrieved passages, citation support and final answer correctness separately. Keep the model and prompt fixed while swapping search providers. Then test the chosen provider at your expected concurrency and context budget.
Benchmark method and source data
Search accuracy 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 →
Common questions
What is the best web search API for RAG?
Best web search API for RAG by accuracy: Perplexity, 77.3% at $5.00/1,000 queries. Fastest: Parallel turbo, 333ms. Cheapest: TinyFish, $0.00/1,000 queries within free-tier limits.
Which benchmark does the main ranking use?
Search accuracy 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.
What does the price per 1,000 queries include?
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.
Does this benchmark measure my complete RAG pipeline?
The main ranking measures search accuracy with a fixed agent using retrieved sources to solve developer questions. Use it to shortlist retrieval providers, then measure final answer correctness and citation support on your own RAG questions.
Related search API comparisons
Web search APIs for AI & LLM developers →Web search APIs & MCPs for AI agents and developers →Search tools for AI agents →Search providers for LLM applications →AI search engines for agents →Free web search APIs for AI agents →Independent web search API comparison →Best fast web search API →Best search and scrape API →Best scrape API for AI agents →Best search API for deep research agents →Best search API for company research →Best search API for sales agents →Best search API for coding documentation →Best web search API for grounding →Best search API for news →Best web search APIs for AI →LLM apps comparison →Pricing per 1,000 queries →