psinglap/websearch-api-eval

Web Search API Eval

A live, side-by-side benchmark of four web search APIs for AI agents: You.com, Tavily, Exa and Parallel. It covers 30 agent-style queries across three use cases (GTM, insurance, finance), with 3 runs per query (90 calls per API), run on October 8, 2026.

• Full results and charts: https://psinglap.github.io/websearch-api-eval/
• Code, harness and methodology: https://github.com/psinglap/websearch-api-eval
• Use it from Claude, ChatGPT or Cursor (MCP): https://websearch-api-eval.psberkeleymba.workers.dev/mcp or npx -y websearch-api-eval-mcp

Subsets

Config	Rows	What each row is
questions (default)	30	One agent query: the upstream task, the exact search call sent to every API, its parameters (freshness window, domain filters), what the agent needs back, the answer key captured live, and source URLs
results	120	One query × provider: a short summary of what the API returned and its accuracy score (0–2, mean of 3 runs, SimpleQA rubric)
scorecards	136	One metric × provider × scope (overall, gtm, insurance, finance): latency, accuracy, freshness, empty responses, content, tokens and cost, as published

from datasets import load_dataset

questions = load_dataset("psinglap/websearch-api-eval", "questions", split="train")
results = load_dataset("psinglap/websearch-api-eval", "results", split="train")

Headline results (overall, 90 calls per API)

Metric	You.com	Tavily	Exa	Parallel
Accuracy score (% of max)	63%	51%	53%	67%
Queries fully correct in all 3 repeats	57%	37%	50%	57%
Results dated inside the query's window	59%	37%	44%	29%
Empty responses (of 90)	0	27	12	3
Latency p50 / p95, client (ms)	593 / 804	1,864 / 3,305	1,106 / 3,997	2,038 / 4,117
Search cost per correct answer	$0.026	$0.016	$0.012	$0.008

Methodology

• Queries. Each query is written the way an agent calls web search: an upstream task, the search call, its parameters and what the agent needs back. Every use case has at least 4 real-time queries; the rest use 24-hour, 48-hour, 7-day or 1-year windows. Some queries expect "none" as the correct answer.
• Same request for every API. 10 results, full content where supported, the freshness window, and domain filters where supported.
• Runs. promptfoo, 3 runs per query. Latency is client wall-clock around the HTTP request after a discarded warm-up call; server-reported latency is kept alongside.
• Grading. Answers were graded from each provider's saved top-5 results with the published SimpleQA grader rubric (correct = 2, partial list = 1, incorrect or not attempted = 0). "None" answers and stale results follow FreshQA strict rules. Answer keys were captured at run time and archived to the Wayback Machine.
• Full metric definitions: eval/METRICS.md.

Limits

• 10 queries per use case is a focused test, not a leaderboard; differences of a query or two are within noise.
• Results reflect one run day (Oct 8, 2026). Freshness and accuracy depend on what was published that day, so answer keys for real-time queries are only valid for that date.
• Costs are vendor list prices on the run date and cover search only.
• returned is a short human-written summary, not the raw API response.

Citation

@misc{singla2026websearchapieval,
  title  = {Web Search API Eval: You.com, Tavily, Exa and Parallel on 30 agent queries},
  author = {Singla, Pallavi},
  year   = {2026},
  url    = {https://github.com/psinglap/websearch-api-eval}
}

License

MIT. Product names and trademarks belong to their owners. This benchmark is independent and was not sponsored by any provider.