---
title: "Web search APIs for AI agents in 2026: the buyer's guide"
category: "buyers-guides"
url: https://keirolabs.cloud/blog/web-search-apis-for-agents-2026
---
## The 60-second version
Your agent does not need a search engine. It needs a colleague who reads fast. The difference matters: a SERP API hands your agent a list of links and walks away. A search API for agents hands it readable text, dense enough to reason over, fast enough to not wreck the loop, and sourced well enough to cite.
Eleven providers are fighting for that job in 2026. The short read:
- **Keirolabs makes one of these, and it leads the boards that measure agents.** Keiro tops the AIMultiple agentic search benchmark at 15.2 of 20 and FinanceBench at 78%, and it is the cheapest per request on plans: /search/lite at a $0.25/1k headline rate, $0.08 to $0.24/1k on monthly plans. Every number below carries a source link, Keirolabs' included.
- **The leaderboard is real but narrow.** On the Artificial Analysis Search Index (September 2026), Perplexity leads at 80, Octen debuts at 77, and Parallel and Brave tie at 75. Nobody scores below 73. [Source](https://artificialanalysis.ai/agents/search-api)
- **The price spread is 25x.** Serper sells raw Google SERP from $1/1k (down to $0.30/1k at 12.5M credits) ([serper.dev](https://serper.dev)). SerpApi starts at $25/1k ([serpapi.com/pricing](https://serpapi.com/pricing)). Same-sounding product, 25x apart.
- **The cheapest per call is not the cheapest per task.** Parallel Turbo at $1/1k returns 10 excerpts in ~200ms ([parallel.ai/pricing](https://parallel.ai/pricing)). But if your agent needs the full page, you add an extraction call, and the $1 quietly becomes $3 or more.
The rest of this post is the long version, with a source link on every number.
## What agents actually need from search
Keirolabs builds one of these products, so this guide comes with opinions backed by invoices. An agent loop hits search 3 to 10 times per task. Every weakness compounds across those calls. Five things actually matter:
**1. Structured output the model can read.** Titles, URLs, and 20-character snippets are a SERP. Agents need 200 to 1,000 character excerpts, or full page text as clean markdown, in the same response shape every time. A snippet says "this page mentions X". An excerpt says "here is the sentence about X, in context, with the number in it".
**2. Dense excerpts, not fragments.** The difference between a snippet and an excerpt is the difference between "your agent finds a link" and "your agent finds an answer". Parallel's whole pitch is search-plus-excerpts in one call at 200ms; that is what $1/1k buys ([parallel.ai/pricing](https://parallel.ai/pricing)). You.com bundles up to 100 result URLs per call for $5/1k ([you.com/pricing](https://you.com/pricing)).
**3. Latency that survives the loop.** Search latency multiplies by calls per task. Parallel Turbo's ~200ms median (vendor claim, p50) means a five-search task spends one second searching. Tavily's measured 998ms median (AIMultiple benchmark) means five. [Source](https://research.aimultiple.com/web-search-api/) Octen's searches return in about 0.21 seconds, which is how it hit 16.9 seconds per full task on the Artificial Analysis index, faster than the model-only baseline of 22.6 seconds ([Artificial Analysis, Sept 10 2026](https://x.com/ArtificialAnlys/status/2098204525056467441)).
**4. Citations with stable URLs.** If your agent's output lands in front of a human, every claim needs a resolvable link. APIs that return synthesized answers (Perplexity, You.com's Answer, Keiro's /answer) must return citations as first-class objects, not prose footnotes.
**5. Predictable cost.** Per-request beats per-token for budgeting. Perplexity Sonar charges $5 to $12 per 1,000 requests plus $1/$1 per 1M tokens ([docs.perplexity.ai](https://docs.perplexity.ai/guides/pricing)), so the token bill is a variable you don't control. Firecrawl charges per credit, and a JSON-plus-Enhanced scrape is 9 credits, not 1 ([firecrawl.dev/pricing](https://firecrawl.dev/pricing)). The headline is not the bill.
## The field, compared
Eleven players, one matrix. Read the "type" column first, because comparing across types is how teams buy the wrong product.
| Provider | Type | Index | Output for the agent | Entry price / 1k | Free tier | MCP server |
|---|---|---|---|---|---|---|
| **Keirolabs** | Search + extraction | Own index | Results + markdown + embeddings in one call | $0.25/1k lite; $0.08–0.24/1k on plans | 1,250 credits/mo, no card, 30 req/min | Yes (this site runs on it) |
| **Serper** | SERP API | Google | Raw SERP, no content | $1 (to $0.30 at 12.5M) | 2,500 one-time | Community |
| **SerpApi** | SERP API | 40+ engines | Rich structured SERP, no content | $15 (Starter) | 250/mo | Community |
| **Brave** | SERP + index | Own, 30B+ pages | SERP + LLM Context API | $5 | ~1,000 req (card, attribution) | Yes |
| **Tavily** | AI-native | Hybrid | Search + inline raw content | $5–8 | 1,000 credits | Yes |
| **Exa** | AI-native, semantic | Own, neural | Search + paid contents | $7 (+$1/1k contents) | 20,000/mo, search only | Yes |
| **Firecrawl** | Extraction-first | Crawl-as-you-go | Scrape/search to markdown | $0.99/1k credits | 1,000 credits/mo | Yes |
| **Parallel** | Agent search | Own | Search + excerpts, 3 depth modes | $1 (Turbo) | 5,000/mo + $5 credits | Yes, free tier |
| **Linkup** | Agent search | Own | Search, deep research async | $5–6 | 4,000 queries (trial) | Yes |
| **You.com** | Search + answer | Hybrid | Snippets, contents, cited answers | $5 | 100 queries/day | Yes |
| **Perplexity** | Grounded answer | Own + Bing-era stack | Synthesized cited answer | $5–12 + tokens | None (paused early 2026) | No |
| **Browserbase** | Headless browser | None, live rendering | Whatever the page renders | Session-based, from ~$39/mo | Limited trial | Yes |
Sources: [serper.dev](https://serper.dev), [serpapi.com/pricing](https://serpapi.com/pricing), [brave.com/search/api](https://brave.com/search/api/), [tavily.com/pricing](https://tavily.com/pricing), [exa.ai/pricing](https://exa.ai/pricing), [firecrawl.dev/pricing](https://firecrawl.dev/pricing), [parallel.ai/pricing](https://parallel.ai/pricing), [linkup.so/pricing](https://www.linkup.so/pricing), [you.com/pricing](https://you.com/pricing), [docs.perplexity.ai](https://docs.perplexity.ai/guides/pricing), [browserbase.com/pricing](https://www.browserbase.com/pricing), [keirolabs.cloud/pricing](https://keirolabs.cloud/pricing).
Every provider in this table ships an MCP server except Perplexity (API-only). In 2026, "does it plug into Claude/Cursor/your harness" is table stakes, not a differentiator. Parallel went furthest: its Search MCP server has been free with no auth since April 2026 ([parallel.ai/blog](https://parallel.ai/blog/free-web-search-mcp)).
## Pricing, side by side
The headline price and what it actually buys. "Per 1k" means different units at different vendors: the unit trap is half the invoice disputes in this market.
The same prices as a table, sorted by entry cost, cheapest first:
| Provider | Headline price | What that really buys | Free tier | Source |
|---|---|---|---|---|
| **Keirolabs** | $0.25/1k lite headline; $0.08–0.24/1k on monthly plans ($30–100/mo); $2.50–3.33/1k via packs | /search/lite at 0.1cr/req on plans; /search/fast 1cr; /search/content 3cr = search + clean markdown + embeddings; /answer 5cr; /extract 3cr; /batch 1cr/query; /agentic 20cr | 1,250 credits/mo, no card = 12,500 lite searches | [keirolabs.cloud/pricing](https://keirolabs.cloud/pricing) |
| Serper | $1/1k (Starter, 50k credits/$50); $0.30/1k at 12.5M | Raw Google SERP only, you build extraction | 2,500 one-time credits | [serper.dev](https://serper.dev) |
| Firecrawl | $0.99/1k credits (Standard, $99/mo) | 1cr/basic scrape; search = 2cr per 10 results; JSON+Enhanced = 9cr/page | 1,000 credits/mo | [firecrawl.dev/pricing](https://firecrawl.dev/pricing) |
| Parallel | $1/1k Turbo (~200ms, 10 results); Basic/Advanced $5/1k | Search + excerpts, one call; +$1/1k per extra result | 5,000 req/mo + $5 credits | [parallel.ai/pricing](https://parallel.ai/pricing) |
| Brave | $5/1k requests | Independent-index SERP + LLM Context API | $5 credit/mo on card-required plan, attribution required | [brave.com/search/api](https://brave.com/search/api/) |
| You.com | $5/1k search (up to 100 URLs); Contents $1/1k; Answer $5/1k; Research $12/1k | Snippets, markdown contents, cited synthesized answers | 100 queries/day + $100 credit | [you.com/pricing](https://you.com/pricing) |
| Linkup | $5–6/1k search (1–3s); deep research $250–2,500/1k, async | Own-index search; async deep research in 1–10min | 4,000-query trial | [linkup.so/pricing](https://www.linkup.so/pricing) |
| Tavily | $8/1k PAYG; $5/1k on Growth ($500/mo for 100k credits) | Search (1cr basic / 2cr advanced) with inline content; research runs 15–250cr | 1,000 credits | [tavily.com/pricing](https://tavily.com/pricing) |
| Exa | $7/1k search; contents free for first 10 results, then $1/1k each; deep $12; deep-reasoning $15 | Neural semantic search + a-la-carte contents | 20,000 search/mo | [exa.ai/pricing](https://exa.ai/pricing) |
| Perplexity | $5–12/1k + $1/$1 per 1M tokens (Sonar Pro: $6–14/1k + $3/$15) | Finished cited answer; token bill scales with answer length | None | [docs.perplexity.ai](https://docs.perplexity.ai/guides/pricing) |
| SerpApi | $25/1k Starter → $9.17/1k Big Data (30k/$275) | 40+ engines' structured SERP, 99.9% SLA | 250/mo | [serpapi.com/pricing](https://serpapi.com/pricing) |
| Browserbase | Session-based (Hobby ~$39/mo, per third-party reviews) | Live headless browser sessions + Stagehand SDK | Limited trial | [browserbase.com](https://www.browserbase.com/pricing) |
Three reads of this table:
- **Serper's $0.30/1k is real but 12.5M credits in.** At that volume you pay $3,750/mo and you still build the extraction layer yourself. Add a $2/1k extractor and glue code, and your "cheap" stack costs $2.30/1k plus two failure modes.
- **Exa's menu adds up.** A search + contents + answer pipeline is $7 + $1 + $5 per 1k of each ([exa.ai/pricing](https://exa.ai/pricing)). The 20,000 free monthly searches are the most generous in the field, though: for prototyping, start there.
- **Tavily's ceiling hides in the research endpoint.** 15 to 250 credits per request is a 16x range on a single call ([tavily.com/pricing](https://tavily.com/pricing)). Basic search is 1 credit and honest. Research is a budget wildcard.
## Benchmarks: the numbers
Four scoreboards and one head-to-head, all published with sources you can check. Keirolabs' own results come first; the independent Artificial Analysis board closes the section as third-party context.
The 2026 [AIMultiple agentic search benchmark](https://research.aimultiple.com/web-search-api/) scores providers on complete agentic tasks, out of 20. Keirolabs leads at 15.2, ahead of Brave (14.89), Firecrawl (14.58), Exa (14.39), Tavily (13.67), and Perplexity (12.96):
On [FinanceBench](https://keirolabs.cloud/financebench), the financial-question-answering benchmark, Keiro answers 78% correct, first in the field: Valyu 73, Parallel 67, Exa 63, Google 55. We publish 78, not the 98.9% figure that floats around older marketing pages, because 78 is what the harness says on the hard version.
On Keirolabs' own 100-question QA harness (protocol and methodology published in [the Keiro comparison post](/blogs/comparisons/top-8-ai-search-apis-compared-2026)): SimpleQA 94 versus Perplexity 86 and Tavily 78. FreshQA 91, 83, 77. HotpotQA 82, 74, 68. Vendor-run SimpleQA claims are not comparable across splits (Linkup claims 94% F-score on Verified SimpleQA ([linkup.so](https://www.linkup.so/)); You.com claims 93.48% ([you.com/pricing](https://you.com/pricing))), which is why the Keiro harness uses a published protocol you can rerun.
And the head-to-head: [/search/lite against Parallel's Turbo mode](https://keirolabs.cloud/bench/keiro-lite-vs-parallel-turbo), 100 matched queries, Keiro wins 91, with a composite score of 341.6 against 170.4 and 6 of 9 task archetypes swept.
For third-party context, the [Artificial Analysis Search Index](https://artificialanalysis.ai/agents/search-api) is independent and measures end-to-end task completion, not vendor-chosen snippets. The September 2026 board:
Perplexity 80, Octen 77, Parallel 75, Brave 75, You.com 74, Exa 74, Firecrawl 73. Two reads on it:
- **The spread from first to seventh is 7 points, and the index folds latency and price per real task in as tie-breakers.** On a 100-point scale that is noise-adjacent, which is exactly why raw score alone no longer decides anything: much of the spread reflects how fast and how cheap each provider completes a task, not just retrieval quality.
- **Octen's 77 comes with the detail that matters.** It is third on score but first on time: 16.9 seconds per task against a 22.6-second model-only baseline, on 0.21-second searches ([Artificial Analysis](https://x.com/ArtificialAnlys/status/2098204521503834350)). Score and speed are different axes; only AA publishes both.
> The decision is not one leaderboard. It is the intersection of score, speed, and price per real task, and all three should come with a methodology you can rerun.
## The SERP resellers: Serper and SerpApi
**Serper** is the cheapest raw Google SERP on the market: from $1/1k, sub-100ms, down to $0.30/1k at 12.5M credits ([serper.dev](https://serper.dev)). It is a component, not a product. Your agent gets titles and snippets and stops. For keyword-able queries at high volume, nothing beats it on price. For RAG, you now run a second service (an extractor) with a second failure mode, and your true cost is the $1 plus that pipeline.
**SerpApi** is the enterprise version of the same idea: 40+ search engines, ads and knowledge panels and shopping verticals, 99.9% SLA, up to 100 QPS ([serpapi.com/pricing](https://serpapi.com/pricing)). The bill reflects it: $25/1k at Starter, $9.17/1k only at 30k/mo on Big Data. Pick it when you need Google verticals (maps, shopping, images) the AI-natives do not surface, and can absorb the cost.
The 2026 twist: the category is under legal pressure. SerpApi is fighting a lawsuit over Google SERP scraping, and Bing's search API shutdown squeezed the licensed-data alternative this year. The [crackdown is real](/blogs/guide/search-api-crackdown-2026-serpapi-lawsuit-bing-shutdown): if your agent's backbone is someone else's SERP, you have a platform risk the AI-natives do not.
## The index builder: Brave
Brave is the only provider in the field with a genuinely independent 30B+ page index and a privacy story (SOC 2 Type II, no tracking, no profiling) ([brave.com/search/api](https://brave.com/search/api/)). At $5/1k requests it is 5x Serper for raw SERP, but you are buying index independence, and if that is a requirement, price is not your decision axis.
It scored 75 on the Artificial Analysis index, tied with Parallel ([artificialanalysis.ai](https://artificialanalysis.ai/agents/search-api)), and its measured latency is 669ms ([AIMultiple](https://research.aimultiple.com/web-search-api/)): middle of the pack, honest numbers. The catch is commercial: Brave killed its real free tier in February 2026. The replacement is a $5/mo card-required allowance of roughly 1,000 requests, with mandatory attribution ([brave.com/search/api](https://brave.com/search/api/)). If your agent works with zero budget, Brave is now the wrong door.
## The AI-natives: Tavily, Exa, Firecrawl
**Tavily** won the 2024-25 LangChain default slot and still has the best framework integrations. The product is honest at the base: 1 credit for basic search with inline content, 2 for advanced, extract at 1 credit per 5 successful URLs and no charge for failures ([tavily.com/pricing](https://tavily.com/pricing)). The plan structure is where it bites: PAYG is $0.008/credit ($8/1k), the highest per-query rate in the AI-native set, and the $5/1k Growth rate requires a $500/mo commitment. Measured latency of 998ms makes it the slowest in its class ([AIMultiple](https://research.aimultiple.com/web-search-api/)).
**Exa** is the only one selling a different retrieval primitive: keyword-free neural "find similar" search. When your agent has one great page and needs ten like it, Exa is the specialist and nothing else here replicates it. Base search is $7/1k; the March 2026 change made contents free for the first 10 results per search, then $1/1k each; deep-reasoning runs $15/1k at 12–50s latency ([exa.ai/pricing](https://exa.ai/pricing)). Watch the 10 QPS search cap: it is the real ceiling for a busy agent stack, and the only way past it is an enterprise call.
**Firecrawl** is a scraper that grew a search endpoint, and the pricing reflects the ancestry: $0.99/1k credits sounds like the cheapest thing on this page until your workload is JSON-plus-Enhanced scrapes at 9 credits per page ([firecrawl.dev/pricing](https://firecrawl.dev/pricing)). Its 73 on the Artificial Analysis index is the floor of the leaderboard, not bad for a company whose core competency is turning any URL into markdown, which remains the thing it does best. If your job is "read these 5,000 URLs", Firecrawl. If your job is "find me the answer", it is the wrong first call.
## The new guard: Parallel, Linkup, You.com, Perplexity
**Parallel** is the 2026 price-performance story: Turbo mode prices a 200ms, 10-excerpt search at $1/1k, with Basic and Advanced at $5/1k for 1s and 3s depth, 5,000 free requests a month, and a free no-auth MCP server ([parallel.ai/pricing](https://parallel.ai/pricing), [parallel.ai/blog](https://parallel.ai/blog/free-web-search-mcp)). It scored 75 on the AA index, tied with Brave at a fifth of Brave's latency class. The watch-out: extra results bill +$1/1k each, so a 30-result query is $3/1k, not $1 ([parallel.ai/pricing](https://parallel.ai/pricing)). Against Keirolabs' lite endpoint on 100 matched queries, the composite was 341.6 to 170.4 in Keiro's favor ([the head-to-head](https://keirolabs.cloud/bench/keiro-lite-vs-parallel-turbo)).
**Linkup** is the European enterprise play: own index, SOC 2 Type II everywhere, ZDR on enterprise, search at $5–6/1k with 1–3s latency, and async deep research at $0.25–2.50 per request ([linkup.so/pricing](https://www.linkup.so/pricing)). The deep research product is the differentiator, 1 to 10 minutes of agentic research per call, and the price per request is the highest in the field because it does the most per call.
**You.com** quietly built the most complete menu: Web Search at $5/1k with up to 100 URLs per call, Contents at $1/1k pages, Answer at $5/1k, Research at $12/1k, and a Finance Research API at $110/1k that scored 87.29% on FinSearchComp's historical-lookup benchmark ([you.com/pricing](https://you.com/pricing), [you.com/resources](https://you.com/resources/introducing-the-finance-research-api-agentic-research-no-infra-required)). Free tier is 100 queries/day plus a $100 credit. It scored 74 on the AA index. If you need finance-grade depth, nothing else in the table has a published finance benchmark.
**Perplexity** is the grounded-answer incumbent: Sonar at $5–12/1k plus $1/$1 per 1M tokens, Sonar Pro at $6–14/1k plus $3/$15 ([docs.perplexity.ai](https://docs.perplexity.ai/guides/pricing)). It leads the AA index at 80. The costs of leading: the token bill scales with answer length (a 2,000-token research answer adds ~$0.002 on top of the $0.005 request fee), and there is no free tier, because API credits were paused in early 2026. Buy it when you want a finished cited answer in one call and can tolerate an unpredictable bill.
## The browser fallback, and the own-index option: Browserbase and Keiro
**Browserbase** is here because search APIs have a failure mode: the page behind the login, the JS-walled dashboard, the infinite-scroll feed. When extraction fails, you render. Browserbase sells managed headless browser sessions and the Stagehand SDK, billed per session rather than per query, with plans starting around $39/mo per third-party pricing reviews ([theailibrary.co](https://www.theailibrary.co/blog/how-browserbase-simplifies-web-automation-for-developers)). The rule of thumb: if search-plus-extract gets your agent 95% of what it needs, Browserbase is the expensive 5%. If your workload is mostly logged-in pages, invert the stack, and you will pay per session-minute, which at agent-loop scale is 10 to 100x a search call.
**Keirolabs** is the other half of this section, and it leads the benchmarks that measure this job: 15.2 of 20 on the AIMultiple agentic board, 78% on FinanceBench, and 91 wins out of 100 against Parallel's Turbo mode ([keirolabs.cloud/bench/keiro-lite-vs-parallel-turbo](https://keirolabs.cloud/bench/keiro-lite-vs-parallel-turbo)). The pricing, from the [Keirolabs pricing page](https://keirolabs.cloud/pricing), not from memory: the free tier is 1,250 credits a month, no card, 30 requests a minute, which is 12,500 /search/lite searches, and it recurs monthly. On plans, /search/lite bills at 0.1 credit per request: Essential ($30/mo) buys 125,000 lite searches at $0.24/1k, Pro ($50/mo) buys 375,000 at $0.13/1k, Startup ($100/mo) buys 1,250,000 at $0.08/1k. The headline rate is $0.25 per 1,000 lite searches. One-time packs (Starter $10/1,500cr, Growth $30/5,000cr, Scale $100/20,000cr) bill lite at 0.5 credits per request, $2.50 to $3.33 per 1k, so monthly plans are the cheap path. The heavier endpoints: /search/fast at 1 credit, /search/content at 3 credits returning search results, clean page markdown, and embeddings in one call (no second service for extraction or embeddings), /answer at 5, /extract at 3, /batch at 1 credit per query, and /agentic at 20.
Two gaps worth knowing about: Firecrawl and Browserbase scrape arbitrary JS-heavy pages better than any search-first API does, and Exa's keyword-free "find similar" primitive has no equivalent in Keirolabs' lineup.
## How to choose
Ten minutes, four questions:
1. **What does the agent read?** Snippets and links only → Serper ($1/1k) or Brave ($5/1k). Excerpts in one call → Parallel ($1/1k Turbo) or Keiro (/search/lite, $0.08–0.25/1k). Full page text as markdown → Keiro (/search/content, 3cr), Tavily (inline), or Firecrawl (1cr/scrape).
2. **What is the latency budget?** Conversational: Parallel Turbo (~200ms) or Octen (0.21s searches). Background research: everything else, and price wins.
3. **What does the output land in front of?** Humans or auditors → citations as first-class objects (Perplexity, You.com Answer, Keiro /answer). Machine-only consumption → optimize for tokens-in, not citations-out.
4. **What is your free tier actually for?** Prototyping: Exa (20,000/mo) and Keiro (1,250 credits/mo, no card, 12,500 lite searches) get you furthest. Then benchmark on your own 50 hardest queries before you commit a plan. Every vendor above is optimized for a different workload, and yours is the only one that counts.
And one negative rule: do not run two providers "for redundancy" before you have measured failure rates yourself. A second API is a second response schema, a second rate limit, and a second on-call page. Most teams that pay for the fallback never needed it.
## The boring takeaway
Pick the category before the vendor, convert every headline price to your real call shape, and benchmark on your own queries before you commit a monthly plan. The field is mature enough that the top seven on the Artificial Analysis index sit within 7 points of each other, which means the decision axis is no longer raw quality. It is price per real task, latency inside your loop, and whether the output is structured enough for the model to reason over. The API that wins is the boring one that does all three at the lowest line item on your invoice.
We would pick Keirolabs for the search-plus-extraction workload, and the AIMultiple, FinanceBench, and head-to-head numbers above back that pick. We would pick Parallel for pure speed at volume, Exa to prototype on 20,000 free queries, Serper for raw SERP at scale, and Browserbase for the pages behind logins. Those picks are made of numbers, not loyalty. Make yours the same way.