What is the best web search API for AI agents in 2026?

On Artificial Analysis's independent Search Index (19 October 2026), Exa and Perplexity score highest at 80, ahead of Octen at 77 and Parallel and Brave at 75. Most list prices run from $0.25 to $25 per 1,000 searches, so choose by your own questions, budget and data terms.

Last checked: 20 October 2026. Disclosure: Grounder (grounder.net) publishes this page. Grounder routes searches to 10 providers (Ozone, Tavily, Exa, Brave, Perplexity, You.com, Firecrawl, Octen, Parallel and Valyu) and is also listed here as a router, so it has a commercial interest in this topic.

Quick comparison

Prices are list prices in US dollars per 1,000 searches, taken from each vendor's own pages. "AA score" is the Artificial Analysis Search Index (0 to 100). A dash means not applicable or not on that index; "not published" means the vendor does not state it.

AA rows were measured between 17 August and 19 October 2026, so ranks are a snapshot. Every AA row uses the same answer model, GPT-5.6 Luna at medium effort. Head-to-head pages: Tavily vs Exa, Tavily vs Brave, Exa vs Brave, Perplexity vs Tavily. Sources: 1, 2

Pick by situation

• Highest independent score: Exa (auto, 80.3) or Perplexity (medium, 80.0).
• Lowest list price: Cloudflare's Web Search API with Ceramic.ai ($0.25, open beta), Serper ($0.30 to $1.00), Parallel fast ($1) or Octen ($1, promotional).
• Lowest search cost per task on AA, among the options covered here: Parallel fast ($8.41 per 1,000 tasks, score 73.1), Octen ($9.07, score 77.1), Firecrawl ($30.48) and Exa lite ($31.20). AA also scores products this guide does not cover, such as TinyFish at $0. Shortest time per task: Octen (14.9 s), Parallel fast (17.6 s), You.com (18.6 s).
• Free to start with no card: Tavily, Exa, Firecrawl, Serper, and Grounder's first 1,000 searches.
• Own index, vendor-stated: Brave, Exa, Parallel, Perplexity. Serper and SerpApi return scraped Google results.
• Licensed or specialist data: You.com, Valyu, Ozone LIVE and Grounder; see licensed content search APIs.
• A cited answer, not a result list: the model-native tools from OpenAI, Anthropic, Google and Microsoft.

AI search APIs

These APIs return ranked results with page text sized for language models. Where a vendor says it runs its own index, that is the vendor's statement. "Grounder run" means Grounder's benchmark, described under "Where Grounder fits".

Exa

• What it is: a search API on Exa's own index, with contents, deep search, answer and agent APIs.
• Pricing: $7 per 1,000 requests (auto, fast), $4 (lite); 10 results included, extra results $1 per 1,000. Free: $10 credit each month, no card.
• Strengths: top of AA (auto with dynamic highlights 80.3; lite 79.2 at $31.20 search cost per 1,000 tasks against $53.17). SOC 2 Type II. Respects robots.txt.
• Limits: the 80.3 needs Dynamic Highlights, a preview with a beta header; auto without it scored 73.7. 10 queries per second by default. Zero data retention is enterprise only. Grounder run: 88.4% SimpleQA, 31.5% SealQA, 5.2k tokens a question.
• Pick it when: you want the highest independent score and accept a preview feature. See Exa alternatives.

Sources: 1, 3, 4, 5, 6, 7

Perplexity

• What it is: a Search API returning ranked results with no generated answer. Sonar Chat Completions ended on 27 September 2026 and moved into the Agent API.
• Pricing: $5 per 1,000 requests; $1 with search_type fast. No free tier is published.
• Strengths: AA 80.0 (medium). 50 queries per second at every tier. SOC 2 Type II. Grounder run: 96.0% SimpleQA (tied first among single providers) at 1.5k tokens and 275 ms; 35.6% SealQA.
• Limits: its privacy page says zero data retention is not the default, while its FAQ says the API keeps no query data. Perplexity-User generally ignores robots.txt.
• Pick it when: you want a raw result list with a high score and can do without a free tier. See Perplexity alternatives.

Sources: 1, 8, 9, 10, 11, 12, 13, 120

Octen

• What it is: a real-time search API with Web, Broad, News, Image and Video search and an Extract API. Public beta since 15 March 2026.
• Pricing: Web Search $1 per 1,000 calls, promotional against a $5 list price with no end date stated. $5 free balance.
• Strengths: AA 77.1 with the lowest search cost of the rows scoring above 75 ($9.07 per 1,000 tasks) and the shortest time per task on AA (14.9 s). 99.9% uptime SLA on paid QPS plans. Grounder run: 96.0% SimpleQA at 276 ms.
• Limits: public beta; no crawler or robots.txt policy published; zero data retention is enterprise only. Paid QPS plans cost $2,099 to $33,999 a month. Grounder run: 21.8% SealQA.
• Pick it when: cost and speed matter most and a beta product is acceptable.

Sources: 1, 14, 15, 16, 17, 18

Parallel

• What it is: an API suite on a proprietary index: Search in four modes, Extract, and asynchronous research APIs.
• Pricing: Search turbo and fast $1 per 1,000 requests; basic and advanced $5; 10 results included. Free: $5 credit a month with a card on file; the Search MCP needs no account.
• Strengths: a complete public price list and limits (600 searches a minute). AA advanced 74.8, fast 73.1. Grounder run: 44.9% SealQA, higher than every other row, and 94.8% SimpleQA.
• Limits: Parallel's changelog says it ranked first on every dimension of AA's index on 19 August 2026; in AA's data of 19 October, advanced (74.8) scores below Exa, Perplexity and Octen. Turbo covers English and Japanese only. No self-serve zero data retention; robots.txt compliance not stated.
• Pick it when: you want low-priced fast modes, a keyless MCP server, or strength on hard multi-source questions.

Sources: 1, 19, 20, 21, 22, 23, 24

Brave Search API

• What it is: a search API on Brave's own index, with an LLM Context endpoint that returns extracted page chunks, and a separate Answers endpoint.
• Pricing: Search plan $5 per 1,000 requests (web, LLM Context, news, images). Answers $4 per 1,000 plus $5 per million tokens. $5 free credit a month; card required; Brave's blog says the credit needs attribution.
• Strengths: AA 74.6 for LLM Context against 64.6 for the Web endpoint. 50 queries per second published. SOC 2 Type II. Grounder run: fewest tokens a question (1.1k on SimpleQA).
• Limits: terms bar caching, storing and AI training on results unless you buy a storage-rights plan with no public price. No hosted MCP server. Grounder run: 84.4% SimpleQA, 16.2% SealQA.
• Pick it when: you want an independent index, low token use and a published rate limit. See Brave alternatives.

Sources: 1, 25, 26, 27, 28, 29, 30

You.com

• What it is: a web and news search API with optional full-page extraction and licensed "knowledge" data, plus Contents, Answer and Research APIs.
• Pricing: $5 per 1,000 calls for up to 100 results, news and knowledge included; extraction adds $1 per 1,000 pages. Free: keyless MCP allows 100 queries a day, and its pricing page shows a $100 free credit (20,000 searches).
• Strengths: AA 74.1 at 18.6 s a task. Honors robots.txt. Licensed feeds (finance, weather, sports) at no extra charge. Grounder run: 84.2% SimpleQA, 20.4% SealQA.
• Limits: zero data retention is enterprise only and excludes Contents and Research. Self-serve limit 10 requests per second.
• Pick it when: you want web results plus licensed structured data under one key.

Sources: 1, 31, 32, 33, 34, 35, 36

Firecrawl

• What it is: a web data API. Search can return results with scraped page content or query-relevant excerpts in one call, beside scrape, crawl, map and parse endpoints.
• Pricing: search costs 2 credits per 10 results; scraping a result adds 1 credit per page. Free: 1,000 credits a month, no card. Hobby ($16 a month, 5,000 credits) to Scale ($599, 1,000,000), billed yearly, is about $6.40 to $1.20 per 1,000 searches (derived).
• Strengths: AA 73.5 at $30.48 search cost per 1,000 tasks (Exa auto: $53.17). Keyless tier. In Claude's connectors directory. SOC 2 Type II. Grounder run: 88.2% SimpleQA at 1.3k tokens.
• Limits: the slowest time per task on AA (60.1 s). Zero data retention is enterprise only and costs five times the credits. No published training-use policy for API content. Grounder run: 19.0% SealQA.
• Pick it when: you want search and page scraping in one call and one credit pool.

Sources: 1, 37, 38, 39, 40, 41, 42

Linkup

• What it is: a search API with three output types (results, sourced answer, structured), plus Fetch and asynchronous Research. Cloudflare's Web Search API also offers it as one of three providers.
• Pricing: $0.005 per request ($5 per 1,000) for results at the flash, fast and standard depths, $0.006 for a sourced answer or structured output, and $0.05 to $0.055 at deep. Free: $20 credit with a professional email, topped up monthly.
• Strengths: a published per-request price list; pay without an account over x402; states it respects robots.txt. SOC 2 Type II.
• Limits: AA 68.3 at a mid-range depth, below Octen, You.com and Firecrawl. 10 queries per second per organization. MCP takes an API key only. Its pages describe zero data retention three ways.
• Pick it when: you want sourced answers or structured JSON at a flat per-request price.

Sources: 1, 43, 44, 45, 46, 47

Tavily

• What it is: Search, Extract, Crawl, Map and Research APIs on one credit balance. Nebius reported the purchase as completed on 19 February 2026; Tavily continues under its own name.
• Pricing: pay as you go at $0.008 per credit: basic search 1 credit ($8 per 1,000), advanced 2 ($16). Project plan $30 a month for 4,000 credits. Free: 1,000 credits a month, no card.
• Strengths: 1,000 requests a minute on paid plans. OAuth hosted MCP, in Claude's directory. Documented integrations with LangChain, LlamaIndex, the Vercel AI SDK and several cloud marketplaces. Grounder run: 85.8% SimpleQA, 24.1% SealQA.
• Limits: AA lists Tavily Search (basic) at 65.6, measured 17 August 2026, with a search cost of $126.44 per 1,000 tasks. Its privacy policy lets it use parts of query data to improve responses and says queries may go to third-party index providers such as Google. No zero data retention documented.
• Pick it when: you want search, extract and crawl behind one key. See Tavily alternatives and migrating from Tavily.

Sources: 1, 48, 49, 50, 51, 52, 53

Valyu

• What it is: a search API over the web and licensed or specialist datasets (SEC filings, papers, clinical trials, patents, market data), with Contents, Answer and DeepResearch APIs.
• Pricing: one credit is $1. Search is billed per 1,000 retrievals: open databases $0.50, web $1.50, financial data $8, proprietary $30 to $50 (docs; its marketing page says $0.50 to $30). Free: $10 credit, $20 with a work email.
• Strengths: licensed Wiley content and public datasets in one API; a stated 24 to 72 hour deletion window; honors robots.txt.
• Limits: no independent benchmark. Grounder run: 63.9% SimpleQA, 13.4% SealQA, 5.9 s median latency. No numeric rate limits published.
• Pick it when: you need filings, biomedical or patent data from the same API as web search.

Sources: 54, 55, 56, 57, 58

Other AI search APIs

• Kagi (public preview since 21 May 2026): $12 per 1,000 searches, Extract $4 per 1,000 pages; lenses and domain rules per request. No free tier or rate limits published; terms bar caching and storing results. Pick it for Kagi's ranking controls.
• Jina (Elastic-owned): Reader turns a URL into markdown; Search returns five results with page text. Billed in tokens (price not published; Search starts at 10,000 tokens a request); 10M free tokens per key. No zero data retention stated. Pick it for URL to markdown.
• Microsoft Web IQ (released 2 June 2026): passages from the Bing index, $7 per 1,000 for web, news, images and video, $1 for browse; no free tier; a paid Azure subscription is required. No SLA; excluded from Microsoft's data protection addendum; its benchmark is Microsoft's own. Pick it if you run on Azure.

Sources: 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69

SERP APIs

SERP APIs return a consumer search engine's results page as JSON. They add no index of their own.

• Serper: Google results; prepaid packs from $50 for 50,000 queries ($1.00 per 1,000) to $3,750 for 12,500,000 ($0.30); credits valid 6 months; 2,500 free queries. Its docs and pricing URLs returned 404 when checked; no official MCP server; no retention or training statement. Pick it for the lowest per-query price on Google results.
• SerpApi: many engines including Google, Bing and Google AI Mode. Free 250 searches a month; Starter $25 for 1,000 ($25.00 per 1,000); Production $150 for 15,000 ($10.00); Searcher $725 for 100,000 ($7.25); Cloud 2M $5,675 for 2,000,000 ($2.84). Official MCP server; search data kept 31 days. Google sued SerpApi under the DMCA on 19 December 2025; on 20 July 2026 the court dismissed both DMCA claims, one without leave to amend and one with, and Google filed an amended complaint on 10 August 2026 (later status not verified). Pick it for several engines in one API.
• Google Programmable Search and Bing Search APIs: Google's Custom Search JSON API is closed to new customers and ends on 1 January 2027; existing customers pay $5 per 1,000 after 100 free a day. Microsoft's lifecycle notice set 11 August 2025 as the retirement date for the Bing Search APIs and named Grounding with Bing Search as the replacement. For the full web, Google's page points to a contact form, and Search Engine Journal reports a partner-only API; no price is published. Use them only to maintain an existing integration. See Google Custom Search alternatives and Bing Search API alternatives.

Sources: 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80

Model-native search tools

These tools run inside a model call. The model decides when to search and writes a cited answer, and you do not control the ranking. Google's and Microsoft's terms also limit caching and storing the results.

OpenAI web search

• What it is: the web_search tool in the Responses API.
• Pricing: $10 per 1,000 calls plus the tokens of the search content. No free tier.
• Strengths: AA 74 (medium) at $40.19 search cost per 1,000 tasks. Domain allow and block lists (up to 100 each), a list of consulted sources, and zero data retention eligibility on the Responses API, with limits.
• Limits: search context capped at 128k tokens; live web access is not HIPAA eligible; OpenAI does not name its index provider.
• Pick it when: you already build on OpenAI models and want the least integration work.

Sources: 1, 81, 82, 83

Claude web search

• What it is: a server tool in the Claude Messages API; Anthropic runs the searches and returns cited sources. A separate web fetch tool retrieves URLs and PDFs at no charge beyond tokens.
• Pricing: $10 per 1,000 searches plus tokens. No free tier.
• Strengths: citations always on; organization-level domain controls; the basic version is eligible for zero data retention.
• Limits: versions with dynamic filtering are not eligible by default; not on Amazon Bedrock; the API docs do not name the search provider.
• Pick it when: you build on Claude and want a flat per-search price.

Sources: 84, 85, 86, 87

Gemini API and Google Cloud grounding

• What it is: Grounding with Google Search, a tool in the Gemini API. Google Cloud also offers Web Grounding for Enterprise, which does not store customer data.
• Pricing: $14 per 1,000 search queries for Gemini 3 and later, 5,000 a month included on the paid tier. Older models: $35 per 1,000 grounded prompts after 1,500 free a day. One request can trigger several billed queries.
• Strengths: Google's index; combines with URL context and code execution.
• Limits: terms require showing Search Suggestions and bar caching, syndicating, indexing or training on grounded results. Log retention is 30 days in the Gemini API and up to 3 days in Google Cloud's Search grounding.
• Pick it when: you build on Gemini or Google Cloud and show answers directly to end users.

Sources: 88, 89, 90, 91, 92, 93

Microsoft Grounding with Bing Search

• What it is: Bing-backed web grounding for Microsoft Foundry agents, including a Web Search tool.
• Pricing: $14 per 1,000 transactions (150 per second, 1 million a day), plus model tokens.
• Strengths: managed inside Foundry with Entra and role-based access.
• Limits: a cited answer, not raw results. Microsoft states that data leaves the Azure compliance boundary and its data protection addendum does not apply. Terms bar storing output and training models on it.
• Pick it when: your agents already run in Foundry.

Sources: 94, 95, 96, 97

Routers

A router puts one API in front of several search providers. For when this helps, see search router vs single provider. Grounder is covered under "Where Grounder fits".

OpenRouter web search

• What it is: web grounding inside OpenRouter's LLM API, through a web plugin (or :online suffix) or a beta server tool. Engines: native, Exa, Parallel, Perplexity or Firecrawl.
• Pricing: Exa $7 per 1,000 requests, Parallel $1 to $5, Perplexity $5, Firecrawl $5 plus $12.50 per 1,000 results; native search at the model provider's rate. Model tokens are extra, even on free models.
• Strengths: one key and one citation format across engines; engine and mode set per request.
• Limits: results come back as model-cited text, not a result list. Server tools are in beta. Retention follows each engine's policy; the EU in-region endpoint has no web search engine.
• Pick it when: you already buy models through OpenRouter.

Sources: 98, 99, 100, 101

Cloudflare Web Search API

• What it is: one endpoint, in open beta since 2 October 2026, that sends a query through Cloudflare's AI Gateway to Ceramic.ai (default), Exa or Linkup, as you choose on each call.
• Pricing: each provider's list price with no markup: Ceramic.ai $0.25 per 1,000 requests, Linkup $5, Exa $7. A 5% fee applies to credits bought through Unified Billing.
• Strengths: the lowest list price in this guide; search logged beside model traffic.
• Limits: beta, with no published rate limits or SLA; up to 10 results per call; no domain or date filters. Cloudflare's pages disagree on whether Exa supports zero data retention through the service.
• Pick it when: you use AI Gateway and want to swap among these three providers with one parameter.

Sources: 102, 103, 104, 105, 106

GroundRoute

• What it is: a separate product from groundroute.ai, not Grounder. One search tool over Serper, Brave Search, Exa, Tavily, Firecrawl and Perplexity.
• Pricing: no subscription or minimum. With your own keys you pay GroundRoute $0 on cache misses and share cache savings 50/50; the managed plan passes through engine cost on misses with the same share.
• Strengths: three cache tiers; automatic fallback to the next engine when one degrades.
• Limits: it routes among general web engines on cost and a quality bar. Its pages cite a 170-query benchmark of its own; the pages we read list no licensed-content source or retention terms.
• Pick it when: your main goal is a lower search bill.

Sources: 107, 108

Licensed content

Ozone LIVE

• What it is: a retrieval-only API over news content Ozone says it licenses from publishers; every row carries the publisher and a licensed flag. Ozone is one of Grounder's 10 providers.
• Pricing: not published; pricing, terms and privacy pages sit behind a sign-in.
• Strengths: a documented OpenAPI spec, a licensed-domain lookup and a Tavily-compatible endpoint.
• Limits: access looks restricted (the sign-in page asks for an @ozoneproject.com address). News only, with partial date coverage on long windows. Its llms.txt states 22,455,219 passages across 1,518 publisher domains (20 October 2026); the vendor's other pages give other figures. Alone, Ozone scored 4.0% SimpleQA and 10.6% SealQA in Grounder's benchmark.
• Pick it when: you need licensed news passages and can get access.

Sources: 109, 110, 111

Data terms at a glance

Sources: 6, 11, 13, 16, 22, 24, 28, 34, 35, 40, 45, 51, 57, 112, 113, 114. For more, see search APIs that respect robots.txt.

Where Grounder fits

Grounder (grounder.net) is a web search API and MCP server for AI agents. It sends each query to the provider, or pair of providers, most likely to answer it well, merges and deduplicates the results, and marks each result with the provider that found it and whether it is licensed. Its 10 search providers are Ozone, Tavily, Exa, Brave, Perplexity, You.com, Firecrawl, Octen, Parallel and Valyu: Ozone serves licensed publisher content, and the other nine search the open web. A search can be pinned to one provider, sent to up to 10 named providers and merged, and capped on time, tokens, provider calls and sources. 18 specialized APIs, such as Polymarket, ESPN, FlightAware and USGS, answer questions about right now.

• Connect: add https://api.grounder.net/mcp to an MCP client and sign in with Google, GitHub or Apple. In Claude Code: claude mcp add --transport http grounder https://api.grounder.net/mcp. Steps for each client: the MCP server guide.
• Free tier: your first 1,000 searches are free, up to 100 a day. No card needed. 120 requests per minute per key.
• Licensed results by default: every signed-up account returns only licensed results (Ozone's publisher content and pages on sites Ozone licenses) unless it turns that off. A search can return fewer results than requested.
• Not included: no fetch, extract, crawl or answer generation. New search keys are no longer made, so new accounts connect over MCP; the HTTP API and drop-in endpoints in Tavily's and Brave's response shapes accept only keys made earlier.
• Data: search records are kept indefinitely and there is no zero data retention. No model is trained on your data.
• Added latency: across 816 searches recorded on 7 and 8 October 2026, each routed and then sent to all 10 providers, a median of 210 ms on top of the provider calls (half between 128 and 266 ms).

Benchmark. Grounder ran this benchmark itself, so read it as a vendor result. It uses 887 questions: 671 from SimpleQA and 216 from the Seal-Hard set of SealQA. From 8 to 13 October 2026 each question was searched once on every provider's default search, taking the top 10 results within 60 seconds. GLM-5.3-Flash at high reasoning effort answered each question from those results alone, and answers were graded against the reference answers with OpenAI's SimpleQA grader. "Grounder optimized" is each question routed as Grounder routes it on 19 October 2026, with licensed sources only, the default. "Grounder optimized, all sources" is the same routing with nothing filtered out. That routing was tuned on these same searches. Table as of 20 October 2026; the data is at GET https://api.grounder.net/v1/demo/results.

Grounder optimized answers only from licensed sources, so it gets fewer answers right than the all-sources row, which searches every source.

On SimpleQA the all-sources row is first (97.8% against 96.0% for Octen and Perplexity). On SealQA it is second to Parallel (36.1% against 44.9%) and just ahead of Perplexity (35.6%). The default row, which every account gets, trades correct answers for licensing: every provider except Ozone scored higher than it on both benchmarks, and against the all-sources row only Parallel on SealQA scored higher. A single provider is the better choice when your questions look like the ones it leads on, when you need fetch or extract, or when you want no router in the path. Grounder fits when you want one MCP address for several providers, the provider shown on every result, licensed results by default or an open benchmark you can re-read. More: About Grounder, Grounder's benchmark.

Sources: 114, 115, 116, 117, 118, 119

Grounder and robots.txt

Allowed sources only is off by default. When it is on, Grounder returns a result only from a site whose robots.txt allows the provider that found it and does not opt out of search or AI answers. A result that cannot be checked in time is left out, so a search can return fewer results. The response then reports, for each provider that answered, the share of its results that were allowed. The check reads each site's robots.txt against the provider's crawler name; it does not show what a provider did when it crawled.

Sources: 114

How we compared

We compared on price per 1,000 searches and what is billed extra, free tier, public benchmark accuracy, latency, output format, index ownership, filters, MCP and integrations, data terms and extras such as extract and crawl. Prices, limits and data terms come from each vendor's own pricing, documentation, privacy and changelog pages, checked on 20 October 2026. Derived numbers (Firecrawl's price per 1,000 searches, Tavily's advanced price) follow from published credit costs. Where a vendor's pages disagree, such as Perplexity's on zero data retention, we say so.

The independent score is Artificial Analysis's Search Index ↗: the equal-weighted mean of DeepSearchQA F1, BrowseComp and AA-Omniscience, with the same answer model, harness and settings for every provider. Vendor-run claims are not used for ranking. The exception is Grounder's own benchmark, which Grounder ran and which is described above; it uses a different protocol, so its scores are not comparable with AA's. Neither tests your questions: run 25 to 100 of your own through two or three options with the same model and prompt, and compare cost per correct answer.

Sources: 1, 2

FAQ

What is the best search API for AI agents?

No single option leads on every measure. On Artificial Analysis's Search Index (19 October 2026), Exa and Perplexity score highest at 80, and of the rows scoring above 75, Octen (77) has the lowest search cost per task. Test two or three on your own questions before you commit.

Which web search API is cheapest?

By list price per 1,000 searches: Cloudflare's Web Search API with Ceramic.ai at $0.25 (open beta), Serper at $0.30 to $1.00 (Google results), then Parallel's fast mode and Octen at $1. Cost per task also depends on queries per task: among the options Artificial Analysis scores, search cost runs from $8.41 per 1,000 tasks (Parallel fast) and $9.07 (Octen) to $126.44 (Tavily basic).

Which search API is fastest?

In Artificial Analysis's data, Octen has the shortest time per task (14.9 s), then Parallel's fast mode (17.6 s) and You.com (18.6 s). In Grounder's benchmark, median search latency was 275 ms for Perplexity and 276 ms for Octen, against 1.9 s for Tavily and 2.1 s for Exa.

Do I need a dedicated search API or is my model's built-in search enough?

Built-in tools take the least work: OpenAI's web search tool costs $10 per 1,000 calls plus tokens and scores 74 on Artificial Analysis's index, which scores a model with no search at 33. The model controls the query and the answer, and Google's and Microsoft's terms restrict caching and reuse of grounded results, so a dedicated API suits agents that need raw results, filters or storage.

Which search APIs have a free tier?

Monthly: Tavily 1,000 credits, Exa $10, Brave $5, Parallel $5, Firecrawl 1,000 credits and Linkup $20. One-time: Octen $5, Valyu $10 to $20, You.com $100, Jina 10M tokens, Serper 2,500 queries and Grounder 1,000 searches, up to 100 a day. Perplexity and Kagi publish no free tier.

Which search APIs have an MCP server?

Tavily, Exa, Perplexity, Parallel, Firecrawl, You.com, Linkup, Valyu, Octen, Jina, SerpApi, Kagi, Microsoft Web IQ and Grounder run hosted servers; Brave's is local only. The built-in search tools from OpenAI, Anthropic, Google and Microsoft are API features, not MCP servers. Add steps are in the MCP server guide.

Should I use more than one search provider?

One is enough when your questions are narrow and you have tested it on them. Rankings change with the benchmark: Exa leads Artificial Analysis's index, while Octen and Perplexity lead Grounder's SimpleQA run and Parallel leads its SealQA run. See search router vs single provider.

Which search APIs offer zero data retention?

Exa, Brave, You.com, Firecrawl, Valyu and Octen offer it on enterprise plans, Linkup on request, and Perplexity and Parallel by agreement. Tavily documents none. Grounder offers none and keeps search records indefinitely.

Next step

Grounder's first 1,000 searches are free, up to 100 a day, no card needed. Connect over MCP at https://api.grounder.net/mcp or read the Docs.