Most Token-Efficient Web Search API for AI Agents, 2026
Most Token-Efficient Web Search API for AI Agents: TinyFish leads 11 web search APIs at 208 points per 1,000 tokens, independently benchmarked Sep 26, 2026.
What this page compares. This page ranks measured results from company-news factual lookup (web search APIs).
How to read the result. The ranking divides the score by tokens returned, so shorter results that still answer rank higher. 2 configurations below 80% of the best score are left out, so nothing ranks on speed or price alone.
Company-news factual lookup, web search APIs, ranked by accuracy per 1,000 tokens
Every row ran the same 300 questions. Provider names link to the official docs for the configuration measured.
| Rank | Provider | Accuracy | Answer recall @5 | Median search time | $ / 1k queries | $ / 1k correct | Result tokens | Accuracy / 1k tokens | List price |
|---|---|---|---|---|---|---|---|---|---|
| 1 | TinyFishfree · 30 req/min cap | 92.0% | 90.7% | 2.3 s | $0 | $0 | 441 | 208 | $0 / 1k queries |
| 2 | Perplexity lowsearch_context_size=low | 97.3% | 98.3% | 1.6 s | $5.00 | $5.14 | 476 | 205 | $5.00 / 1k queries |
| 3 | SERPlimit=10 | 96.0% | 95.0% | 578 ms | $3.00 | $3.13 | 497 | 193 | $3.00 / 1k queries |
| 4 | Firecrawlfirecrawl-search | 95.3% | 96.7% | 471 ms | $5.00 | $5.24 | 678 | 141 | $5.00 / 1k queries |
| 5 | Brave Search webcount=10 · result_filter=web | 93.3% | 91.7% | 613 ms | $5.00 | $5.36 | 817 | 114 | $5.00 / 1k queries |
| 6 | Tavily basicsearch_depth=basic | 87.7% | 87.7% | 1.7 s | $8.00 | $9.13 | 1,639 | 53.5 | $8.00 / 1k queries |
| 7 | Exa fasttype=fast | 99.3% | 99.3% | 569 ms | $7.00 | $7.05 | 1,987 | 50.0 | $7.00 / 1k queries |
| 8 | Parallel fastmode=fast | 86.0% | 79.0% | 861 ms | $1.00 | $1.16 | 1,839 | 46.8 | $1.00 / 1k queries |
| 9 | Exa instanttype=instant | 97.7% | 97.3% | 386 ms | $7.00 | $7.17 | 2,128 | 45.9 | $7.00 / 1k queries |
| 10 | Brave Search llm-contextcount=10 | 94.0% | 94.7% | 590 ms | $5.00 | $5.32 | 2,064 | 45.6 | $5.00 / 1k queries |
| 11 | Tavily advancedsearch_depth=advanced | 93.0% | 92.7% | 4.2 s | $16.00 | $17.20 | 2,210 | 42.1 | $16.00 / 1k queries |
| 12 | Parallel basicmode=basic | 93.3% | 92.0% | 1.8 s | $5.00 | $5.36 | 2,330 | 40.1 | $5.00 / 1k queries |
| 13 | Nimble standardsearch_depth=standard · full_content=false · focus=general | 93.0% | 94.3% | 822 ms | $5.00 | $5.38 | 2,730 | 34.1 | $5.00 / 1k queries |
| 14 | You highlights coreextraction_mode=highlights · knowledge=core | 92.0% | 89.7% | 815 ms | $5.00 | $5.43 | 2,862 | 32.1 | $5.00 / 1k queries |
| 15 | Linkup fastdepth=fast · outputType=searchResults | 96.7% | 94.7% | 1.2 s | $5.00 | $5.17 | 3,022 | 32.0 | $5.00 / 1k queries |
| 16 | You highlightsextraction_mode=highlights | 90.7% | 89.7% | 614 ms | $5.00 | $5.51 | 2,837 | 32.0 | $5.00 / 1k queries |
| 17 | Linkup standarddepth=standard · outputType=searchResults | 92.0% | 90.3% | 2.1 s | $5.00 | $5.43 | 2,983 | 30.8 | $5.00 / 1k queries |
How the ranking was measured
- Every API received the same natural-language question about a recent company event, one search per question, up to 10 results.
- Each answer is a single verifiable value pinned to an official wire or newsroom URL, fixed before any vendor ran.
- Accuracy is whether the returned results let the same answering model extract the correct value.
- Latency is the search request time. Cost is the recorded list price for the calls made, per 1,000 queries and per 1,000 correct answers.
- The runner and published data are open in openbenchmarks-labs/factual-lookup-company-news-search ↗ and openbenchmarks/OB-News-Websearch ↗. The full methodology is on the Web Search Benchmark →
Questions answered by this comparison
Which provider leads most Token-Efficient Web Search API for AI Agents?
TinyFish leads this table at 208 points per 1,000 tokens, on 300 company-news questions.
How was this measured?
Every API received the same natural-language question about a recent company event, one search per question, up to 10 results. Each answer is a single verifiable value pinned to an official wire or newsroom URL, fixed before any vendor ran. Accuracy is whether the returned results let the same answering model extract the correct value. Latency is the search request time. Cost is the recorded list price for the calls made, per 1,000 queries and per 1,000 correct answers.
How should speed and cost be read beside the score?
TinyFish recorded 2.3 s median search time and $0 per 1,000 queries. The most accurate configuration is rarely the fastest or the cheapest, so each is reported in its own column.
Do vendors pay to be ranked?
No. Inclusion and rank do not depend on vendor payments. Every API used the same questions, answering model and scoring.
What does this page not test?
This page ranks measured results from company-news factual lookup (web search APIs).