Search APIs for AI agents: every vendor wins its own benchmark

When an AI agent "searches the web," it almost always queries a paid search API like Brave, Exa, Tavily or, more recently, DeepSeek. We compared the public benchmarks: vendors win their own tests, and in independent tests the leaders are usually within the margin of error. Plus an overview of pricing, speed and privacy.

Analysis Nov 15, 2026 10 min read Original analysis by our editors
Illustrative image, AI-generated

What this is about

When a homegrown assistant, a coding agent or a ChatGPT clone "looks something up online," there's almost never a proprietary search engine behind it. The agent sends the query to a search API – an interface that returns search results as text for a fee. The best-known providers are Brave, Exa, Parallel, Tavily and Firecrawl. Then there are services like SerpApi and cloro that automatically scrape Google's results page, and since November, DeepSeek with a budget search offering.

Which of them delivers the best answers is currently answered mostly by the vendors themselves – each in their own favor. We compared the public benchmarks and checked who ran the tests, with which model, and how big the error bars are. We didn't run our own test for this article.

Why this matters right now

The market has shifted significantly in just a few weeks:

  • DeepSeek has opened up the web search and page reading from its app as an API. Since November 1 (Beijing time), Search and Fetch each cost $0.75 per 1,000 calls, according to its pricing page. According to ppc.land and eesel.ai, only successful requests are billed. Each account is limited to 5 requests per second, and search queries can be at most 128 characters long.
  • On November 13, Cloudflare released its agent browser Kitesurf under the Apache 2.0 license (github.com/cloudflare/kitesurf). It's not a search engine but a lightweight browser engine that lets agents open and read the pages they find.
  • In November, new comparison articles appeared from Parallel, Tavily and Firecrawl, plus an update from AIMultiple on November 10.
  • In the background, the Bing Search API has been gone since August 11, 2025. Microsoft shut it down and points to "Grounding with Bing" in Azure, which according to ppc.land costs $35 per 1,000 transactions. That means there's no longer a second major Western index besides Google that you can simply buy access to.

Vendor tests: eight benchmarks, eight winners

Test Who ran it Winner Notable issues
SealQA/SimpleQA (Aug. 2026) Tavily Tavily (55.7 % SealQA-Hard) number of questions not stated; grader GPT-4.1-mini
SimpleQA, 4,326 questions Firecrawl Firecrawl (94.7 %) grader not described
SimpleQA Verified/BrowseComp/WideSearch (Sept.) Parallel Parallel Advanced (97 % SimpleQA) BrowseComp only 50 questions, tied with Perplexity
SimpleQA, FreshQA, HotpotQA Keiro Keiro (94/91/82) grader is the small Gemma 3 12B
FreshQA, 358 questions (Nov. 5) cloro cloro/SerpApi (88.3/88.5 %) cloro sells one of the two winners
SERP comparison, 50 queries cloro cloro "This ranking is ours," with affiliate links
FreshQA, 500 questions (2025) You.com You.com (F1 0.44) competitor scores only in charts
Overview report Keenable "Parallel leads" auto-generated, now outdated

The pattern is clear: whoever builds the test picks the questions, model, grader and competitor settings – and ends up on top. In a November comparison, Parallel itself writes that the numbers are "still ours," meaning in-house. A small irony: Keenable's overview report still lists Parallel as the leader. On the current Artificial Analysis leaderboard, Keenable itself now sits at the top.

Independent tests: the leaders are within the noise

Artificial Analysis Search Index. This is the most rigorous controlled test. It comprises 1,700 tasks from three benchmarks (900 DeepSearchQA, 200 BrowseComp, 600 AA-Omniscience). All providers run with the same model (GPT-5.6 Luna) and the same agent harness, which can search and open pages for up to 25 rounds. Only the search API changes. Without search, the model scores 33 points. As of November 9, Keenable pro (83) and Keenable realtime (82) lead, followed by Exa auto and Perplexity medium at 80 each. Parallel advanced and Brave score 75, DeepSeek Search 69. At launch in August, according to The Decoder, Parallel was still in front with 75, and in September it was Perplexity. The methodology doesn't describe confidence intervals or repeated runs. The grader is the same model that produces the answers.

If you add up search and model costs per 1,000 tasks, the picture shifts:

Provider (AA Index) Score Search + model per 1,000 tasks
Keenable pro 83 $40.22
Exa auto 80 $63.77
Perplexity medium 80 $70.34
Octen 77 $23.57
Brave LLM Context 75 $78.91
DeepSeek Search 69 $28.66

At $8.47, DeepSeek has the cheapest search on the list but incurs the highest model costs ($20.19). Our take: the list price per search says little about total cost. Weaker results mean more reading work for the model.

AIMultiple (updated November 10) tested 24 configurations of six APIs with 120 questions. Three judge models graded the answers, and an answer only counts as correct if at least two agree. Parallel Advanced reaches 87.0 percent, but the 95 percent interval runs from 78.3 to 94.2. Exa Fast comes in at 82.6 percent (73.9 to 91.3). All five top configurations overlap, and AIMultiple itself says the difference is uncertain. The freshness section was tested with so few questions that the interval runs from 57.9 to 94.7 percent. On whether the returned snippets actually support the answer, Brave leads at 96.8 percent. Each configuration was asked each question only once.

Vals Web Search Index (finance and law): Keenable with GPT-5.6 Sol scores 49.08 percent, Exa with Claude Fable 5 scores 48.45 percent. That's a gap of 0.6 points. The page lists Exa, Keenable, Parallel and Tavily under "Partners." Vals doesn't explain what that partnership means.

The freshness test: Google scrapers ahead, but measured by a scraper vendor

The widely cited FreshQA test on dev.to is methodologically solid: 358 questions whose answers change over time, a third of them with a false premise. There was one search per question, a fixed answer model (gpt-5.6-luna), a fixed grader (gpt-5.6-sol) and pairwise significance tests. The harness and graded answers are public.

API Correct 95 % interval $ per 1,000 Median latency
SerpApi (Google) 88.5 % 84.8–91.4 7.25 8.20 s
cloro (Google) 88.3 % 84.5–91.2 1.92 7.96 s
Brave 84.9 % 80.8–88.3 5.00 1.21 s
Exa 83.0 % 78.7–86.5 7.00 0.74 s
Parallel 82.7 % 78.4–86.2 5.00 2.99 s
Tavily 81.8 % 77.5–85.5 5.00 2.00 s

But the test was run by cloro, which sells one of the two Google scrapers. The author discloses this. Still, the test isn't independent. There's also a caveat on the substance: Brave trails cloro by a margin that isn't statistically significant (p = 0.065), and there's no difference at all between Brave, Exa, Parallel and Tavily. Much of Google's lead comes from the AI Overview at the top of the results page. Without it, accuracy in a pilot run with 243 questions dropped from 90.5 to 80.2 percent. In other words, the agent is reading an answer that Google's own AI has already written. The test didn't measure how fresh the indexes are. With SerpApi, a second search was also billed in 75 percent of cases.

Is the most accurate source even legal?

Google is suing SerpApi in federal court in California (Case No. 4:25-cv-10826-YGR). The allegation: SerpApi circumvents the "SearchGuard" bot protection and thereby violates the DMCA, the US copyright law that prohibits circumventing technical protection measures. Judge Yvonne Gonzalez Rogers dismissed the suit on July 20. Google filed an amended complaint on August 10, citing among other things its licensing deal with Reddit. On November 9, the court dismissed that version too, because Google hadn't shown which copyrighted content SearchGuard actually protects. Google gets one more chance to amend – according to PPC Land, until around November 30. If it fails, the suit will be dismissed for good. Search Engine Roundtable stresses that the court did not rule that the scraping is legal. In a parallel case brought by Reddit against SerpApi in New York, by contrast, the suit has not been dismissed.

Pricing, sources and privacy at a glance

Provider Source of results $ per 1,000 searches (list) Limits/latency Privacy
Brave own index, according to Brave over 35 billion pages 5 50/s US company; ZDR Enterprise only; usage per DPA only, according to Brave
Exa own crawler + embedding index 4 (lite) to 7 (auto/fast), Deep 12–15 Enterprise 25/s ZDR Enterprise only (annual plan $12,000)
Parallel not disclosed in our sources 1 (turbo/fast), 5 (basic/advanced) 600/min (Build) ZDR from Scale with contract; data location not stated
Tavily own "Tavily Search Index" 8 (pay-as-you-go, basic), advanced double n/a owned by Nebius since 2026; claims GDPR compliance; no ZDR documented
Firecrawl web search: source not disclosed; own specialized indexes approx. 2–10 depending on plan 10–10,000/min ZDR Enterprise only
SerpApi scraped Google pages 25 (Starter) to 7.25 (100,000/month) 8.2 s "ZeroTrace" only from $3,750/month
cloro scraped Google pages 1.92 (100,000/month) 8.0 s not documented in sources
DeepSeek own search engine 0.75 (+0.75 per Fetch) 5/s, 128 characters servers in China; training on inputs is default per policy, opt-out by email

ZDR stands for zero data retention: the provider doesn't store requests in the first place. A DPA is the data processing agreement required under the GDPR when a service provider processes personal data on your behalf. With DeepSeek, there's a caveat: its privacy policy explicitly excludes end-user data from API customers' apps. eesel.ai found no specific retention rules for the search endpoints. A tester there reported that politically sensitive queries return no results. A Hacker News user who runs a fetch benchmark himself called DeepSeek's search "cheap, but not very good."

Our verdict

Our take: there's no clear winner, and whoever declares one usually built the test themselves. In controlled tests with a fixed model, the top group is often separated by just a few points, with interval widths of up to 16 points. The ranking also depends heavily on the task: single fact questions, multi-step research or current events.

In practice, that means for you:

  • Freshness: If you need current events, Google scrapers are the most accurate – but you pay with roughly 8 seconds of latency and unresolved legal risk.
  • Speed: Brave and Exa deliver similarly good results in under a second and a half.
  • Price: DeepSeek is by far the cheapest. But with servers in China and no retention rules, it's hard to justify for queries involving personal data from the EU.
  • Privacy: Real ZDR is almost always available only with an enterprise contract.

In the end, the most reliable benchmark is one you run with your own questions. Several providers offer free tiers that make this easy.

Sources

#search api#ai agents#benchmarks#deepseek#privacy#comparison

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.