Best Web Search API for AI Agents for Deep Research 2026: Independent Benchmark, Open Source
Independent 2026 benchmark (open source code, open data) of 11 web search APIs (Exa, Tavily, Brave, Parallel, Firecrawl): best web search API for AI agents for deep research, ranked on multi-hop F1 (search-only and search & fetch).
Deep research for an AI agent is multi-hop company search. Each question stacks three or four constraints. One query is not enough; the agent has to assemble a complete set from several searches. Returned names are resolved to canonical companies and scored on F1. Search-only and search & fetch stay separate because opening pages changes both quality and cost. The model-only baseline is 0.
Best web search API for AI agents for deep research
Primary rank is F1 on 45 multi-constraint company searches. Pick the ranking that matches whether the agent can fetch.
- Search-only: Parallel basic: 46.5% F1. Snippets, no fetch.
- Search & fetch: Exa deep: 48.2% F1. Agent opens pages.
- Model-only baseline: 0. The model cannot complete these questions without a search API.
Deep research, search-only: F1
Provider search results and snippets. No page-reading tool. 45 questions, three independent agent runs.
| Provider | Endpoint & configuration | F1 | Precision | Recall | Exact set |
|---|---|---|---|---|---|
| Parallel | POST /v1/search mode=basicparallel-basic | 46.5 ± 1.9 | 88.7 ± 0.9 | 34.4 ± 2.0 | 3.7 ± 2.6 |
| Exa | POST /search type=deepexa-deep | 45.4 ± 2.0 | 83.2 ± 3.0 | 33.7 ± 1.4 | 1.5 ± 1.3 |
| Parallel | POST /v1/search mode=advancedparallel-advanced | 44.2 ± 1.4 | 87.6 ± 3.4 | 32.0 ± 1.6 | 2.2 ± 0.0 |
| Exa | POST /search type=instantexa-instant | 43.3 ± 1.0 | 82.6 ± 3.8 | 32.2 ± 0.5 | 3.0 ± 1.3 |
| Linkup | POST /v1/search depth=fastlinkup-fast | 41.1 ± 1.7 | 82.8 ± 1.0 | 30.3 ± 1.5 | 0.7 ± 1.3 |
| Tavily | POST /search search_depth=advancedtavily-advanced | 41.1 ± 2.3 | 83.7 ± 4.2 | 29.8 ± 1.7 | 2.2 ± 0.0 |
| Linkup | POST /v1/search depth=standardlinkup-standard | 40.6 ± 0.9 | 84.0 ± 7.9 | 29.7 ± 1.4 | 1.5 ± 2.6 |
| Parallel | POST /v1/search mode=fastparallel-fast | 38.0 ± 2.0 | 79.9 ± 4.6 | 27.5 ± 1.4 | 1.5 ± 1.3 |
| Perplexity | POST /searchsearch_context_size=low | 37.8 ± 2.1 | 79.3 ± 5.7 | 26.8 ± 1.4 | 2.2 ± 2.2 |
| Parallel | POST /v1/search mode=turboparallel-turbo | 34.7 ± 2.4 | 80.0 ± 5.2 | 24.8 ± 2.4 | 1.5 ± 1.3 |
| You | POST /v1/searchyou | 33.1 ± 2.4 | 75.7 ± 3.1 | 23.1 ± 1.9 | 1.5 ± 1.3 |
| Firecrawl | POST /v2/searchfirecrawl | 30.4 ± 1.1 | 77.3 ± 4.4 | 20.7 ± 0.7 | 2.2 ± 0.0 |
| Brave Search | GET /res/v1/web/searchbrave | 28.0 ± 1.7 | 66.9 ± 8.5 | 19.3 ± 0.5 | 0.7 ± 1.3 |
| TinyFish | GET api.search.tinyfish.aitinyfish | 26.6 ± 1.3 | 64.7 ± 1.5 | 17.9 ± 0.8 | 1.5 ± 1.3 |
| Seltz | POST /v1/search scope=companiesseltz-companies | 14.5 ± 0.9 | 40.0 ± 3.1 | 9.4 ± 0.7 | 0.0 ± 0.0 |
| SERP (RapidAPI) | GET google-search74.p.rapidapi.comserp | 0.4 ± 0.6 | 0.7 ± 1.3 | 0.3 ± 0.4 | 0.0 ± 0.0 |
F1, precision, recall, and exact-set accuracy are percentages reported as mean ± sample SD across three independent runs; each run aggregates all 45 questions. SD is measured in percentage points.
Deep research, search & fetch: F1
Same questions. The agent can open an exact URL returned by search. Separate ranking from search-only.
| Provider | Endpoint & configuration | F1 | Precision | Recall | Exact set |
|---|---|---|---|---|---|
| Exa | POST /search type=deepexa-deep | 48.2 ± 2.1 | 89.4 ± 1.2 | 36.0 ± 2.2 | 2.2 ± 2.2 |
| Perplexity | POST /searchsearch_context_size=high | 46.6 ± 2.0 | 87.7 ± 5.9 | 34.7 ± 1.1 | 2.2 ± 2.2 |
| Exa | POST /search type=instantexa-instant | 44.9 ± 0.9 | 85.9 ± 2.5 | 33.5 ± 0.9 | 5.2 ± 1.3 |
| Parallel | POST /v1/search mode=basicparallel-basic | 42.3 ± 1.1 | 81.3 ± 2.8 | 31.3 ± 1.1 | 3.0 ± 1.3 |
| Parallel | POST /v1/search mode=advancedparallel-advanced | 42.2 ± 1.1 | 87.6 ± 0.3 | 30.1 ± 1.1 | 2.2 ± 0.0 |
| Linkup | POST /v1/search depth=standardlinkup-standard | 42.0 ± 1.8 | 90.7 ± 2.0 | 30.5 ± 2.0 | 3.0 ± 3.4 |
| Tavily | POST /search search_depth=advancedtavily-advanced | 41.0 ± 1.3 | 89.4 ± 6.0 | 29.1 ± 0.7 | 2.2 ± 0.0 |
| Linkup | POST /v1/search depth=fastlinkup-fast | 39.9 ± 1.3 | 85.3 ± 3.6 | 28.6 ± 1.3 | 0.7 ± 1.3 |
| Parallel | POST /v1/search mode=fastparallel-fast | 39.3 ± 3.3 | 82.3 ± 6.6 | 28.2 ± 2.0 | 2.2 ± 0.0 |
| Parallel | POST /v1/search mode=turboparallel-turbo | 36.0 ± 3.5 | 83.6 ± 4.0 | 25.0 ± 2.6 | 0.0 ± 0.0 |
| You | POST /v1/searchyou | 34.0 ± 0.9 | 78.8 ± 0.6 | 23.8 ± 1.2 | 3.0 ± 1.3 |
| Firecrawl | POST /v2/searchfirecrawl | 33.2 ± 2.1 | 83.3 ± 0.8 | 22.7 ± 1.8 | 1.5 ± 1.3 |
| TinyFish | GET api.search.tinyfish.aitinyfish | 30.2 ± 3.5 | 70.9 ± 11.3 | 20.9 ± 2.5 | 0.0 ± 0.0 |
| Brave Search | GET /res/v1/web/searchbrave | 29.4 ± 1.6 | 73.5 ± 5.8 | 20.4 ± 0.9 | 1.5 ± 1.3 |
| Seltz | POST /v1/search scope=companiesseltz-companies | 16.3 ± 1.5 | 49.5 ± 2.4 | 10.2 ± 1.2 | 0.0 ± 0.0 |
| SERP (RapidAPI) | GET google-search74.p.rapidapi.comserp | 0.0 ± 0.0 | 0.0 ± 0.0 | 0.0 ± 0.0 | 0.0 ± 0.0 |
F1, precision, recall, and exact-set accuracy are percentages reported as mean ± sample SD across three independent runs; each run aggregates all 45 questions. SD is measured in percentage points.
How deep research is scored
One model and one prompt across providers (gpt-5.6-sol, medium reasoning). Max 8 model turns and 14 searches. Every returned company is resolved to a canonical identity and compared with a human-reviewed gold set. A correct member is a true positive, an extra company is a false positive, and a missed gold company is a false negative. Precision, recall, and F1 are averaged across questions and three independent agent runs.
Open runner: openbenchmarks-labs/multi-turn-company-search · Open data: openbenchmarks/OB-Company-Websearch
Fact finding and coding agents are separate rankings
Best web search API for AI agents for fact finding 2026: one query, one fact, search-only snippets.
Web search for coding agents: grounded patches on held-out docs, not a company set.
Ranked for one constraint on this job: most accurate · fastest · cheapest
Exa vs Tavily vs Brave
Same F1 rankings as above. Pairwise pages also show lookup and coding for that pair.
Exa vs Tavily · Tavily vs Parallel · Brave Search vs Exa · Linkup vs Firecrawl · Exa alternatives
Best web search API for AI agents for deep research: FAQ
What is the best web search API for AI agents for deep research in 2026?
Deep research here is multi-hop company search: 45 questions with three or four constraints, scored on F1 against a human-reviewed gold set. Search-only leader: Parallel basic at 46.5% F1. Search & fetch leader: Exa deep at 48.2% F1. Open source code and open data. Use fetch if the agent can open pages.
How is deep research different from fact finding?
Fact finding is one query, one fact, search-only snippets. Deep research needs several searches to assemble a complete company set. The model-only baseline on this job is 0: the model cannot answer from training data. Use the fact-finding ranking for one-shot lookup.
Should my deep research agent use search-only or search and fetch?
Search-only ranks snippet results. Search & fetch ranks agents that open pages returned by search. They are different jobs and different APIs can lead. Use search-only numbers if the tool cannot fetch. Use search & fetch if it can.
Why F1 instead of extracted-answer accuracy?
The answer is a set of companies, not one string. Extra companies lower precision. Missed gold companies lower recall. F1 combines both. Exact-set accuracy is the stricter case where the returned set matches the reviewed set exactly. Lookup accuracy does not apply.
Can the model do deep research without a web search API?
No. The model-only baseline is 0. Each question combines three or four constraints that are not conclusively present in training data. A complete company set has to be assembled from several searches.
Exa vs Tavily vs Brave for deep research?
They sit on these same F1 rankings. Pairwise pages (Exa vs Tavily, Tavily vs Parallel, Brave Search vs Exa, Linkup vs Firecrawl) cover deep research plus lookup and coding. This page does not average them into a best overall.
Where is cost and speed for deep research?
This page ranks quality (F1). Time / task quality and LLM $ / task quality are on the fastest and cheapest multi-hop pages, and on the full multi-turn company search benchmark.
Is this benchmark independent, with open source code and open data?
Yes. No vendor pays for inclusion or rank. The question, prompt, model (gpt-5.6-sol, medium reasoning), turn budget (max 8 turns, 14 searches), and output schema are held constant; only the search API changes. Runner: https://github.com/openbenchmarks-labs/multi-turn-company-search. Dataset: https://huggingface.co/datasets/openbenchmarks/OB-Company-Websearch. 45 questions are live on this page.