The Best Web Search API for AI Agents in 2026
Use Exa when your agent needs page text with results, the Perplexity Search API for the lowest cost per query, and the Brave Search API for an independent index at a flat $5 per 1,000 requests. Exa and Brave run their own indexes; Perplexity counts up to five queries as one billed request.
Key takeaways
- Exa runs its own index and returns page text with results: $7 per 1,000 searches, $1 per 1,000 pages, and $10 of free credit every month.
- The Perplexity Search API is $5 per 1,000 requests ($1 with Fast Search) and counts up to five queries as one request.
- Perplexity ended support for Sonar Chat Completions on September 27, 2026; its Agent API is the replacement.
- The Brave Search API uses an independent index at $5 per 1,000 requests, with $5 of free credit a month.
Why does an AI agent need a search API?
A language model only knows what was in its training data, and that data stops months before you use it. Ask about a price, a release or anything that happened last week and it will either refuse or, worse, answer confidently and wrongly.
A search API is the fix. The agent turns the question into a query, the API returns ranked results — and sometimes the page text — and the model answers from those sources with citations. It is the same pattern as retrieval over your own documents, pointed at the whole web.
Current answers
Prices, docs, news and changelogs from this week, not from whenever the model was trained.
Citations users can check
Every claim can link to its source, which is what makes an AI answer trustworthy enough to act on.
Fewer made-up facts
A model that answers from retrieved text has much less room to invent than one answering from memory.
No scraping Google
Google's terms prohibit automated queries, and scrapers break when layouts change. These are licensed APIs with stable formats.
What it does not do: a search API does not make an agent right. It returns whatever the web says, including outdated, wrong and SEO-spam pages, and the model still has to judge them. Search quality also varies by query type, so test on your own questions rather than trusting anyone's benchmark, ours included.
Do you need this yet? If your app is a chatbot on a single model, the provider's built-in web search tool may be enough — OpenAI, Anthropic and Google all offer one. Reach for a dedicated search API when you want to control queries and sources yourself, batch many searches cheaply, reuse results across models, or keep search working if you change model provider.
Quick Verdict
Its own index, ranked by meaning, with page text, highlights or summaries returned in the same call. Deep search and research agents on top.
Raw ranked results with up to five queries per billed request, plus an Agent API that writes cited answers with the model you choose. Sonar Chat Completions has been retired.
An independent web index, not a Google or Bing reseller, at a flat price per request. An Answers plan adds cited AI answers.
TL;DR
No single winner — pick on what your agent needs back. Page text with the results → Exa. The lowest price per query, or a cited answer without writing the loop → Perplexity. An independent index at a flat rate → Brave.
What does a web search API call return?
The price per 1,000 searches is the number everyone compares, and it is the least useful one. What matters is what comes back, because whatever the API does not return, your agent has to fetch, clean or write itself — and pay for.
Links and snippets — Brave, Perplexity Search API
Ranked results with titles, URLs and text snippets. Cheapest per call and fine when snippets are enough to answer, or when you only need to find the right page. For full text, you fetch and clean pages yourself or with a scraping API.
Results plus page contents — Exa
The search can return cleaned page text, highlights or summaries alongside the links, so the model reads sources without a separate scraping step. You pay for the content, at $1 per 1,000 pages per content type.
Finished answers — Perplexity Agent API, Brave Answers, Exa Answer
The provider searches, reads and writes the cited answer for you. The least code, and the least control: you pay search fees plus model tokens, and the cost per question moves with how much the model reads.
Using Perplexity Sonar? Perplexity ended support for Sonar Chat Completions on 27 September 2026. Synchronous and streaming requests are being reformulated as Agent API requests, rolling out by model; asynchronous Sonar requests no longer work and need to move to the Agent API's background mode. Plan the migration rather than waiting for behaviour to shift under you.
How do Exa, Perplexity and Brave compare?
Prices and allowances as published on each vendor's pricing page or docs in October 2026. All three bill by usage with no required subscription.
Exa
Its own index, ranked by meaning
$7 per 1,000 searches
$4 on lite; $10 free credit every month
- Search: $7 per 1,000 (auto, fast), $4 (lite), up to 10 results
- Contents: $1 per 1,000 pages per type (text, highlights, summary)
- Deep Search $12 per 1,000; Answer $5 per 1,000
- Monitors re-run a query on a schedule, $15 per 1,000
- $10 at signup, balance resets to $10 each month
- Enterprise: custom indexes, up to 1,000 results per search
- CaveatContents are billed per page on top of the search, and the agent and deep endpoints scale up fast — a high-effort agent run is $0.50 to $1.00. Neural search also rewards descriptive queries, so keyword-style prompts may need rewriting.
Best for: Research agents and RAG over the web that need page text, or 'find pages like this' discovery.
Perplexity
Raw results, or cited answers via the Agent API
$5 per 1,000 requests
$1 per 1,000 with Fast Search
- Search API: up to five queries in one request, billed once
- Failed and rate-limited requests are not billed
- Agent API web_search: $2.50 per 1,000 calls ($1 Fast)
- fetch_url: $0.50 per 1,000 calls for page content
- Agent API runs Sonar or models from OpenAI, Anthropic, Google, xAI
- Sonar Chat Completions support ended 27 Sep 2026
- CaveatNo free monthly allowance. In the Agent API the bill is search fees plus the chosen model's tokens, so a cheap search can sit inside an expensive answer. Existing Sonar integrations need migrating.
Best for: High-volume agents where cost per query matters, or apps that want a cited answer without building the loop.
Brave Search
An independent index at a flat rate
$5 per 1,000 requests
$5 free credit every month; card required
- Search plan: up to 50 queries per second
- Extra snippets, schema-enriched results, Goggles re-ranking
- Answers plan: $4 per 1,000 + $5 per million tokens
- Answers is OpenAI SDK compatible, with citations
- Prepaid, no subscription
- Enterprise: zero data retention, custom agreements
- CaveatFull page contents are not included, so a reading agent needs a fetcher too. Storing results or training on them requires a plan with storage rights — check before you cache results long-term.
Best for: Teams that want results from an index independent of Google and Bing, at a price that is easy to forecast.
How much do 10,000 searches a month cost?
A research agent handling 10,000 questions a month: one search per question with 10 results, then reading the top three pages — 30,000 pages. List prices, October 2026, pay as you go, after each free allowance.
| API | 10,000 searches | Reading 30,000 pages |
|---|---|---|
| $70 ($60 after the $10 credit); $40 on lite | +$30 for text of 30,000 pages | |
| $50 standard; $10 with Fast Search | Snippets only; fetch pages yourself | |
| $45 after the $5 credit | Snippets only; fetch pages yourself |
Read the second column first: the cheapest search is not the cheapest agent. Brave and Perplexity look lowest until the agent needs page text, at which point a scraping API or your own fetcher joins the bill. Exa costs more per search because it can do the reading too. And if your agent batches related queries, Perplexity's five-per-request rule can cut its search line by up to 80%.
Feature comparison
Other routes worth knowing
Two alternatives cover a lot of agents without a separate search vendor.
Your model provider's web search tool
OpenAI, Anthropic and Google each offer web search as a tool inside their model APIs, billed per search on top of tokens. One vendor, one bill, no glue code — at the cost of control over queries, sources and switching providers later.
A scraping API with a search endpoint that returns pages as clean markdown. A natural partner for Brave or Perplexity when the agent needs full text, or a single-vendor alternative if scraping is most of the job.
Which one should you pick?
Your agent reads pages, not just links
Exa. Search and page text in one call means one vendor, one bill and no scraper to maintain. Use highlights rather than full text where you can — fewer tokens for the model as well.
You run a lot of searches and cost per query rules
Perplexity Search API with Fast Search and batched queries, or Brave if you want an independent index at a flat rate. Pair either with a fetcher for the few pages you actually need to read.
You are not sure which returns better results for you
Take 50 real questions from your users, run them through two contrasting APIs — say Exa and Brave — and grade the answers your agent produces, not the result lists. All three have free allowances or low minimums, so the test costs a few dollars.
Frequently Asked Questions
What is the best web search API for AI agents in 2026?
There is no single best one; they return different things. Exa runs its own index and returns page contents in the same call. The Perplexity Search API is the cheapest per request and can batch up to five queries in one billed call. The Brave Search API gives you an independent index at a flat $5 per 1,000 requests. Choose by whether your agent needs links, page text or a finished answer.
Exa vs Brave Search API: what is the difference?
Both run their own web index, but they return different things. Exa ranks by meaning, so you can describe the kind of page you want, and can return page text, highlights or summaries in the same call; it bills $7 per 1,000 searches plus $1 per 1,000 pages, with $10 of free credit a month. Brave returns classic ranked results with snippets at a flat $5 per 1,000 requests, with $5 of free credit a month, and leaves fetching full pages to you.
What happened to Perplexity Sonar?
Perplexity ended support for Sonar Chat Completions on 27 September 2026 and recommends its Agent API instead. Synchronous and streaming Sonar requests keep working because they are being reformulated as Agent API requests, rolling out by model, but asynchronous Sonar requests are no longer supported. The Agent API lets you choose the model and adds a web_search tool billed at $2.50 per 1,000 calls. For raw results without an answer, the Search API is $5 per 1,000 requests.
Is there a free web search API?
Two of the three have a monthly free allowance. Exa gives $10 of credit that resets on the first of each month. Brave applies $5 of free credit a month, about 1,000 searches, but requires a card. Perplexity’s Search API has no free allowance listed. Both allowances are enough for development and a small production app.
Should I use my model provider’s built-in web search instead?
Often, for a simple chatbot. OpenAI, Anthropic and Google all offer web search as a tool in their model APIs, billed per search on top of tokens. A separate search API makes sense when you want to control the query, the sources and the results yourself, reuse results across models, batch many searches cheaply, or keep search working if you switch model provider.
Can I just scrape Google instead?
Not reliably or legitimately. Google’s terms prohibit automated queries, and scraped results break whenever the page layout changes or your IPs get blocked. Unofficial SERP-scraping APIs exist but carry the same terms risk. The APIs on this page are licensed products with stable formats, which is what you want under an agent that runs unattended.
Sources
Prices were read from each vendor’s pricing page or documentation on October 23, 2026.