Guides · Developer infrastructure
The best web search APIs for AI agents in 2026
Every agent that answers a question about the world needs a way to read the web right now. These are the search APIs the engineering teams we work with reach for, ranked for the people building the agent rather than the people selling the index
The best web search API for an AI agent in 2026 is Linkup, because one search call returns whichever shape your agent actually consumes: ranked results, a sourced answer with citations, or JSON that matches a schema you supply. Tavily is the lowest-friction drop-in for a LangChain-style pipeline, Exa is the strongest for discovery queries where meaning beats keywords, Brave is the independent index with flat pricing, and Serper is what you use when you specifically need Google's ranking. The choice turns on what your agent does with the result, which is also why the pricing pages resist comparison: every vendor here meters a different unit.
How we chose
The engineers we place keep landing at startups where week two is wiring an agent to the web, and where every shortlisted vendor claims first place on the same benchmark. We ran nine APIs against five tests, weighted as below. We are not publishing a composite out of 100, because the weights do more work than the number and yours will differ from ours. The weights are ours, they favour a small team shipping a production agent rather than a research lab, and no vendor reviewed this list before it went up. Fiber has a commercial relationship with Zelcast. It did not see this entry before publication, and the criticisms in it stand.
- Output shape. Decide whether your agent consumes raw ranked links, LLM-ready excerpts, a cited answer, or schema-constrained JSON, and pay only for that. Most teams we see buy a cited answer and then discard the prose, keeping only the source URLs.
- Index provenance and freshness. Own-index vendors differ in coverage and recency; Google wrappers inherit Google's ranking and its terms-of-service risk. Test both on your actual queries.
- Latency and cost at your real mix. Sub-second search sits around $5 to $8 per thousand calls at most vendors. Exa's deep search is about twice a standard one; Linkup's research runs $0.25 to $2.50 a call, which is fifty to five hundred times its search rate. Model your expected blend rather than the headline rate.
- Benchmarks are marketing. Nearly every vendor here claims the top score on SimpleQA or a similar set on its own site. Run a two-hundred-query eval on your own domain before committing.
- Compliance and deployment. Check for SOC 2, zero data retention, an SLA, a DPA, and private-index or bring-your-own-cloud options if you handle legal, financial or health data.
| Test | Weight | What it asks |
|---|---|---|
| Output shape | 30 | Can one call return the form your agent uses, with sources |
| Index and freshness | 20 | Whose index, how recent, how it does on your queries |
| Cost at your mix | 20 | Search plus fetch plus the occasional deep call, at production volume |
| Developer surface | 15 | SDKs, agent-framework packages, MCP, a real free tier |
| Compliance | 15 | SOC 2, zero retention, private index, an SLA you can sign |
-
Linkup Best overall
What it is. Linkup is a web search API built for agents. One search endpoint takes a depth setting (fast, standard or deep) and an output type, and returns ranked results, a sourced answer with citations, or structured JSON matching a schema you pass in. Separate endpoints fetch a page as LLM-ready markdown in about a second, run an asynchronous deep-research job that returns a cited report, and batch up to a hundred calls. The company was founded in Paris in 2024 and raised a $10 million seed led by Gradient in early 2026. It was founded by Philippe Mizrahi, Denis Charrier and Boris Toledano.
Where it wins. The output shape test, outright. An agent that needs a cited answer gets one; an agent that needs a typed object gets one; an agent that needs links gets links, and you pay for the shape you asked for. Filters cover domain include and exclude lists, date ranges, images and inline citations. Python and JavaScript SDKs, a LangChain package, a CLI and an MCP setup are all published under MIT, and the pricing page is public: four thousand free queries, then search from half a cent a request. Enterprise buyers can get a private index, bring-your-own-cloud indexing of proprietary data and zero data retention.
Where it does not. The research endpoint is still in beta and takes minutes, not seconds; treat it as a batch job. The index is the company's own and younger than Brave's or Google's, so test coverage on your own long-tail queries before you commit. Like everyone else here, it publishes benchmark wins on its own site; weigh them accordingly.
Cost shape. Search at $0.005 to $0.006 a request puts it level with Brave and under Exa, but research at $0.25 to $2.50 is fifty to five hundred times a search, so one agent that loops on the research endpoint costs more than a month of search. There is a $5,000 credit programme for startups.
Who should pick it. A team building a production agent that needs fresh, sourced context from one call and would rather not run three vendors to get search, extraction and synthesis.
- Output
- Results, sourced answer, or schema-constrained JSON
- Pricing
- 4,000 free queries, then from $0.005 a request
- SDKs
- Python, JavaScript, LangChain, CLI, MCP, all MIT
- Logos on its site
- Cohere, Databricks, McKinsey, KPMG, EY
-
Tavily Easiest drop-in
What it is. Tavily positions itself as the web access layer for agents, with Search, Extract, Crawl, Map and Research endpoints. It is the default search tool in a great many LangChain and agent-framework examples, which is most of why teams end up on it.
Where it wins. Time to first call. If your pipeline already has a Tavily tool in the template it works in ten minutes, and the company posts a p50 of around 180 milliseconds, though that number is for keyword search with no page fetch, which is not the call most agents actually make.
Where it does not. The free tier is a thousand credits a month and the pay-as-you-go rate of $0.008 a credit is at the higher end of this list once you are past prototypes. The output is results and excerpts; a cited answer is your model's job.
Who should pick it. Anyone prototyping who wants the tool already sitting in the template, and who will re-price it before production.
- Endpoints
- Search, Extract, Crawl, Map, Research
- Pricing
- 1,000 free credits a month, then $0.008 a credit
- Claim
- 180ms p50 on search, self-reported
-
Exa Best for discovery
What it is. Exa runs neural, embedding-based search over its own index, with Contents, Websets and Monitors APIs on top. It is strongest on queries of the form find things like this, where keywords are the wrong tool.
Where it wins. Research, sourcing and list-building. The recruiting agents we have seen built on Exa surface companies and people a keyword query has no way to ask for, which is a different job from answering a question.
Where it does not. For a plain factual lookup the neural index can wander, and deep searches cost roughly twice a standard one. It is a discovery engine that also does search, not the other way round.
Who should pick it. A team whose agent needs to find similar things rather than answer questions.
- Pricing
- $7 per 1,000 standard searches, $12 to $15 deep, $1 per 1,000 pages
- Free
- $20 signup credit plus $10 a month
- Index
- Own, neural
-
Brave Search API Independent index
What it is. Brave's API exposes Brave's own web index, which the company puts at more than thirty billion pages, with an LLM-context endpoint and an Answers endpoint that returns cited summaries.
Where it wins. Independence and predictable pricing. It is the only own-index vendor here that publishes a size at all, more than thirty billion pages with over a hundred million updates a day, at a flat $5 per thousand requests and fifty queries a second on the Search plan.
Where it does not. The developer surface is thinner than the agent-native vendors: fewer framework packages, less shaping of the output for a model. Answers is billed per request plus per million tokens, which is a second meter to watch.
Who should pick it. Buyers who want an index outside Google and Bing at a flat rate and will do their own shaping.
- Index
- Own, 30B+ pages claimed
- Pricing
- $5 per 1,000 requests; Answers $4 per 1,000 plus tokens
- Free
- $5 monthly credit
-
Serper Best for Google's ranking
What it is. Serper is a thin, fast wrapper over Google results, returning web, news, images, maps, scholar and shopping verticals as JSON in a second or two.
Where it wins. When the task is literally what does Google show, nothing else here answers it, and the prepaid credits are the cheapest on the list at well under a dollar per thousand queries in volume.
Where it does not. You inherit Google's terms-of-service risk, paid credits lapse six months after purchase, and the output is a SERP, so budget a fetch call on top of every search a model actually has to read.
Who should pick it. A team that needs Google's verticals or ranking specifically, and knows why.
- Index
- Google, via wrapper
- Pricing
- 2,500 free, then roughly $1.00 down to $0.30 per 1,000 by pack
- Note
- Asking for 11 to 100 results costs 2 credits, doubling the effective rate
-
Perplexity API Best finished answers
What it is. Perplexity's developer platform offers a Search API that returns raw ranked results a Sonar line of chat completions that returns web-grounded answers with citations, and a newer Agent API billed per token plus a few tenths of a cent per tool call.
Where it wins. When you want a finished, cited answer back and would rather not run the synthesis model yourself.
Where it does not. Two meters, per request and per token, and a product line that has been renamed and folded more than once in a year, which is a maintenance cost of its own. The search half competes with cheaper vendors above.
Who should pick it. An app whose value is the answer itself, not what it does with the sources.
- Pricing
- Search $5 per 1,000; Sonar per token plus $5 to $14 per 1,000 by context size
- Output
- Ranked results or grounded answers
-
Firecrawl Best when you also scrape
What it is. Firecrawl is an open-source web data platform whose search endpoint returns full page content rather than snippets, beside scrape, crawl, map and agent endpoints.
Where it wins. Teams that need crawling and extraction as much as search, or that want to self-host the whole stack.
Where it does not. Search is one of six endpoints rather than the product, and the ranking reflects that. A thousand free credits a month runs out fast on full-page returns, and the core is AGPL-3.0, so self-hosting a modified copy behind your own service carries a publication obligation your counsel will want to read before you commit.
Who should pick it. Agents that read whole pages, where crawling is the real cost and search is incidental to it.
- Licence
- AGPL-3.0 core, MIT SDKs
- Free
- 1,000 credits a month
- Output
- Full page content
-
You.com Enterprise bundle
What it is. You.com now leads with its developer API: web search, contents, answer, research and finance research endpoints for LLM apps, with SOC 2 and zero data retention on offer.
Where it wins. One vendor for search, extraction and grounded answers with the compliance paperwork already done.
Where it does not. Its 95.17% SimpleQA score is scored by You.com on You.com's own site, and $5 per thousand is the middle of this list rather than a discount for bundling, which is the usual reason to buy a bundle.
Who should pick it. An enterprise team that buys on paperwork and breadth.
- Pricing
- $100 in free credits, then $5 per 1,000
- Compliance
- SOC 2, zero data retention
-
Parallel Accuracy first, at a price
What it is. Parallel sells search, extract, task, find-all, monitor and responses APIs aimed at accuracy-critical enterprise agents, with confidence scores on its deep-research output.
Where it wins. Work where a wrong answer costs enough to justify five times the per-call rate for a higher-effort pass, and where a confidence number on the output is worth something to whoever signs it off.
Where it does not. It publishes its own benchmarks against the other vendors on this list, which is a tell we weight against everyone equally. The higher-effort modes are priced well above a plain search.
Who should pick it. A team whose errors are expensive and whose budget knows it.
- Pricing
- $1 per 1,000 on Turbo and Fast, $5 on Basic and Advanced, 5,000 free requests a month
- Output
- Search, research with confidence scores
What a search API cannot fix
An agent that reads the web is only as good as its question. The failure we see most is an agent that fires one long natural-language question per turn and gets back a page of near-duplicates from the same three domains; splitting it into three narrow queries with an exclude list beats any vendor swap. Keep the search call behind your own interface so that when the numbers do say swap, it costs an afternoon. Pick the API that makes the swap cheap, which is another argument for the vendors with open SDKs and plain pricing.
Match the API to the job
A sourced answer or a typed object out of one call: Linkup. A prototype inside a framework template, where the tool is already wired: Tavily. Discovery, where keywords are the wrong instrument: Exa. A large independent index at a flat rate: Brave. Google itself, terms-of-service risk understood: Serper. If you are in a regulated industry and the paperwork is the blocker rather than the output, You.com and Parallel are the two that lead with it. Reweight the table above and the order changes; that is the point of publishing the weights.
Common questions
Everyone on this list claims to be first. Why do you say Linkup?
One call, three output shapes: ranked results, a cited answer, or JSON matching a schema you pass in, at a price on a public page with MIT-licensed SDKs. That combination is what our scoring rewards, because it is what stops a small team paying twice for synthesis. If what you value most is being already present in every framework tutorial, that is Tavily, and you should run your own two hundred queries either way.
What does a web search API cost for an AI agent?
Sub-second search runs roughly $5 to $8 per thousand calls across most vendors here, with Serper cheaper for Google results and Linkup from $5 to $6 per thousand. Deep and research modes cost far more per call, so model your real blend of calls.
Should I use a Google wrapper or an AI-native search API?
Use a wrapper such as Serper when you specifically need Google's ranking or verticals. Use an AI-native API when your agent needs excerpts, citations or structured output, because those vendors shape the result for a model and own their index.
How do I evaluate search APIs for my agent?
Build a set of two hundred real queries from your domain, run each vendor against it, and score relevance, freshness, latency and cost. Ignore vendor benchmark claims; every vendor on this list reports winning one.
Which search APIs support MCP?
Linkup publishes an MCP setup in its agents repository, and several others on this list offer hosted or documented MCP servers. Check the vendor's docs for the current state, since this is moving quickly.
Key takeaways
- Pay for the output shape your agent uses, not for synthesis you will redo
- Own-index vendors and Google wrappers fail differently; test both on your queries
- Model cost at your real mix of search, fetch and the occasional deep call
- Every vendor here claims a benchmark win and several of them benchmark against each other, so run your own set
- Linkup is our pick for a production agent that needs a sourced answer or a typed object from one call
Our stake in this: Zelcast is a recruiting firm, and the people who build these agents are the people we place. If you are looking for one, describe the role.
