Best Search APIs for AI Agents and LLMs (2026): Prices and Limits Compared

An LLM on its own knows only what was in its training data. A search API gives an agent or a RAG pipeline current information: it sends a query and gets back results (and often page text) that the model can read and cite.

The options changed in 2025: one major official web search API was retired and another closed to new customers. This guide covers what to look for, compares the main options with prices taken from each vendor's own pricing page, and says honestly who should pick which. If your model writes its own answers, start with Serpex, the search API we make: one request returns up to 10 results and, on request, the page content of the top 5 or 10 as Markdown, at $0.50 to $0.80 per 1,000 searches, with errors and timeouts never billed and 200 free credits to start, no card.

We make one of these APIs, Serpex, so read this as a vendor's guide. Every competitor figure below links to its source and carries the date it was checked, and each API gets a "pick it if" line, including the ones we compete with.

The short answer: which API for which job

What changed: the Bing and Google APIs

If you are reading this because an older integration stopped working, this is probably why.

• Bing Search APIs were retired on August 11, 2025 (Microsoft announcement). If you still have code calling the Bing endpoints, see our Bing Search API alternative page for a migration path.
• Google's Custom Search JSON API is closed to new customers. Existing customers have until January 1, 2027 to transition, and Google points to Vertex AI Search for searching up to 50 domains (Google developer docs). That covers a limited set of domains, not the whole web. Our Google Custom Search API alternative page covers the options if you need web-wide results.

For a new project that needs web-wide results, that leaves the independent search APIs below.

How to choose a search API for an agent

Six questions decide most of it. Answer them before comparing prices.

1. Do you want retrieval or an answer?

Some APIs return results and leave the reasoning to your model. Others, such as Perplexity's Agent API or Tavily's answer option, run the search and write a cited answer with their own model. If you already run a capable model with your own prompts, you probably want retrieval. If you want to remove a model call from the pipeline, an answer engine may suit you better.

2. Does page content come with the results?

Snippets are rarely enough for grounding. Check whether the API can return full page text with the search results, or whether you need a second call to an extract endpoint, with its own price and its own failure cases.

3. What controls does the request expose?

Domain include/exclude lists, date ranges, topic modes and search depth are request parameters on some APIs and missing on others. If your agent must restrict sources, check this first.

4. What exactly is billed?

Compare cost per task, not just cost per call. Check whether errors, empty results and repeated queries are charged, whether content is priced per page or per token, and whether "deep" modes cost a multiple of the base rate.

5. Is there a free tier you can test on?

A renewing monthly allowance is different from a one-time signup credit. Both are fine for evaluation; only one works for a hobby project that runs indefinitely.

6. How does it plug in?

Official SDKs, an MCP server for agent clients, and existing LangChain or LlamaIndex integrations all reduce the code you have to write.

Comparison table

Search prices are per 1,000 requests at list price. Competitor figures are from each vendor's own pricing page, checked 2026-10-23 (Firecrawl and Jina Reader as of 2026-08-16; sources in the sections below). For a wider price table, including Google results page APIs, see Search API Pricing in 2026. Serpex figures are from our pricing page.

Two tools come up in the same conversations but are not search APIs: Firecrawl crawls sites you already know and turns pages into markdown, and Jina Reader converts a URL you already have into markdown. They are covered briefly at the end.

Each API, honestly

Serpex

Serpex is a search API: real-time web search results as structured JSON, plus page content extraction. POST /api/search returns title, url, snippet and position per result; with include_content: true the top 5 or 10 results also carry their page content as markdown. A separate Extract API returns markdown or HTML for up to 10 URLs per request. There are Python and TypeScript SDKs, an MCP server (serpex-mcp), and LangChain, LlamaIndex, Haystack and n8n integrations.

Billing: errors and timeouts are never charged, unconfirmed empty results are free, your own repeated requests within about 5 minutes cost 0 credits, and page content is billed on what was delivered. Content is best-effort; the docs say it arrives for roughly 70% of requested results. Our SLA states a 99% monthly uptime target, with service credits for unplanned downtime caused by Serpex.

Limits: no domain, date or topic filter parameters, no generated answers, no site crawling, no PDF extraction.

Pay as you go is $0.80, $0.65 or $0.50 per 1,000 searches depending on top-up size, and volume plans start at $0.20 per 1,000 for 10M+ searches a month, with custom rate limits.

Pick it if your model does its own reasoning and you need results plus page text at volume, for $0.50 – $0.80 per 1,000 searches. Docs · Pricing

Tavily

A credits-based search API with a broad set of RAG-specific features: answer synthesis, several search depths, domain and time filters, and Crawl, Map and Research endpoints. Its LangChain and LlamaIndex integrations are well established. Pay as you go is $8 per 1,000 basic searches ($16 advanced), and $5 on the $500-a-month Growth plan. Sources: Tavily credits docs, checked 2026-10-23; endpoints and parameters from Tavily's docs, checked 2026-09-28.

Pick it if you want the API to synthesise answers, rely on its filters, or live inside the 1,000 free monthly credits. Compare in detail: Serpex vs Tavily · Tavily alternative

Exa

Exa's embeddings-based search finds pages by meaning rather than keywords, which suits open-ended research queries. It is pure pay-as-you-go, has deep search and answer endpoints, and offers mature filtering (date ranges, domain include/exclude, categories). Contents are a separate call, with text, highlights and summaries billed separately. Search is $4 (Lite), $7 (Auto) or $12 – $15 (Deep) per 1,000 requests. Source: Exa pricing docs, checked 2026-10-23.

Pick it if semantic recall on research-style questions matters more to you than cost per query. Exa alternative · Tavily vs Exa

Parallel

Parallel sells web search and deep research for agents, priced per request, and is the closest to Serpex on search price in this table. Its Task API runs multi-step deep research as a managed job, and processor tiers let you trade cost for depth. Search is $1 per 1,000 in Turbo or Fast mode and $5 in Basic or Advanced mode, with compressed excerpts included and full pages through a separate Extract API. Source: parallel.ai/pricing, checked 2026-10-23.

Pick it if you need managed deep research rather than search primitives. Parallel alternative

Linkup

Linkup has a standard search mode and a deep mode that does much more retrieval work per query ($5 per 1,000 standard searches, $50 per 1,000 deep searches). It has licensed premium content partnerships, and accounts that qualify get their $20 credit topped back up each month. Source: Linkup pricing docs, checked 2026-10-23.

Pick it if you need its licensed premium sources, or your volume fits in the monthly credit. Linkup alternative

Brave Search API

Brave runs its own independent index rather than reselling another engine's results, has a strong public privacy stance with no user profiling, and charges a flat price per call, including for its LLM Context endpoint. Its $5 monthly free credit needs a card on file. Source: brave.com/search/api, checked 2026-10-23.

Pick it if index independence or privacy guarantees are requirements you have to answer for. Brave Search API alternative

Perplexity Search API

Perplexity ended support for its Sonar Chat Completions API on September 27, 2026. For retrieval it now sells a Search API that returns ranked web results with extracted content, with controls for domains, languages and regions, at $5.00 per 1,000 requests ($1.00 with Fast Search). Answer-style use moves to its Agent API, billed per tool call plus model tokens. Source: Perplexity pricing docs, checked 2026-10-23.

Pick it if you want ranked results with extracted content and Perplexity's domain, language and region controls. Perplexity API alternative

You.com

You.com sells a Web Search API at $5.00 per 1,000 calls on its Standard plan, and page content through a separate Contents API at $1 per 1,000 pages. Source: you.com/pricing, checked 2026-10-23.

Pick it if you want search and page contents from one vendor as two separate, predictable line items.

Parallel vs Tavily, and the deep-research APIs

These come up together often enough to answer directly.

Markdown and page text. Tavily returns cleaned page content on the search call when you set include_raw_content. Parallel's search returns compressed excerpts of each page; for the full page you call its Extract API ($1 per 1,000 URLs). Serpex returns Markdown for the top 5 or 10 results in the same call and bills only the pages it delivered. If your agent needs whole pages in one round trip, Tavily or Serpex; if excerpts are enough, Parallel's are built for exactly that.

Price per 1,000 searches. Parallel is $1 in Turbo/Fast mode and $5 in Basic/Advanced. Tavily is $8 pay as you go and $5 on Growth, with advanced depth counting double. Serpex is $0.50 – $0.80, and quoted from $0.20 at 10M+ searches a month.

Deep research. This is a different product from search: you hand over a question and get a researched result back. Parallel's Task API runs it as a managed job with processor tiers, Exa's Deep search costs $12 – $15 per 1,000, Linkup's deep mode $50 per 1,000, and Tavily has a Research endpoint. Choose a research API when you want the vendor to plan the steps; choose search when your own agent plans them.

Not search APIs, but often paired with one

Google results page APIs such as cloro and Serper return the Google page itself as JSON: the organic results and, in cloro's case, the AI Overview and AI assistant answers too. That is a different job from a search API: you get exactly what Google or an AI assistant shows for a query, not page text prepared for your model. Use one when what Google shows is the thing you need to know, for example to track how an AI Overview answers questions in your market.

Firecrawl crawls and extracts whole sites into markdown, with depth and path control, schema-based extraction and first-class JavaScript rendering. It costs $0.75 – $3.80 per 1,000 pages on monthly billing, with 1,000 free credits a month (firecrawl.dev/pricing, as of 2026-08-16). Use it when you know which site you want crawled thoroughly. Firecrawl alternative

Jina Reader turns a URL into markdown by prefixing it, billed on response tokens ($0.050 per million, $0.045 at volume). Its 10 million free tokens are one-time and licensed for non-commercial use only (jina.ai/reader, as of 2026-08-16). Use it when you already have the URLs. Jina Reader alternative

How to evaluate before you commit

Published prices get you to a shortlist. Your own queries decide the rest.

1. Build a query set from real traffic. Take 50 – 100 queries your agent actually issues, including awkward ones: long-tail, recent events, non-English.
2. Run each shortlisted API on the same set within its free tier.
3. Score what your model needs, not what looks good: were the right pages in the results, and did page text actually arrive for them?
4. Compute cost per completed task. Include extract calls, deep-mode multipliers and token fees. Where the API reports credits per response, as Serpex does with metadata.credits_used, sum those rather than estimating.
5. Check the failure path. Look at what comes back when a page cannot be fetched or a query has no results, and whether you were charged for it.

Buying for a company, with contracts, security reviews and rate limits to agree? See Search APIs for Enterprise AI in 2026.

FAQ

What replaced the Bing Search API?

Microsoft retired the Bing Search APIs on August 11, 2025. If you need web-wide results, the independent search APIs in the table above are the options to evaluate. Our Bing Search API alternative page covers migration.

Can I still sign up for Google Custom Search JSON API?

No. It is closed to new customers, and existing customers have until January 1, 2027 to transition. Google suggests Vertex AI Search for searching up to 50 domains. See our Google Custom Search API alternative page.

Which search API is best for a RAG pipeline?

It depends on the answers to the six questions above. If your model writes the answer and you need page text at volume, a retrieval API that returns content with results (Serpex, or Tavily with raw content) keeps the pipeline simple. If you want the API to write the answer, look at Perplexity's Agent API or Tavily's answer option. For open-ended semantic research, look at Exa.

Do I need a separate crawler alongside a search API?

Often not. Several search APIs return page text with the results or offer an extract endpoint. You need a crawler such as Firecrawl when you have to cover an entire known site rather than the pages a search returns.

What are the best Parallel alternatives for AI search?

For plain search at a similar or lower price per query, Serpex ($0.50 – $0.80 per 1,000, quoted from $0.20 at 10M+ a month) and Perplexity's Fast Search ($1). For managed deep research like Parallel's Task API, Exa Deep, Linkup's deep mode or Tavily Research.

Parallel vs Tavily: which is better for Markdown?

Tavily returns cleaned full-page content on the search call with include_raw_content. Parallel returns compressed excerpts and sells full pages through its Extract API. For whole pages in one call, Tavily or Serpex; for compact excerpts, Parallel.

Which search APIs return structured output in real time?

Every search API in the table returns JSON with a title, URL and snippet per result, fetched at request time. The difference is what comes with it: full page text (Serpex, Tavily), extracted content or excerpts (Perplexity, Parallel), or semantic matches (Exa).

How do I add search results to an AI workflow?

Give the model search as a tool: the model writes the query, your code sends it, and the results go back into the prompt with their URLs so the model can cite them. With an API that returns page text, that is one call per search. Our Python SDK shows the call, and the MCP server adds it to agent clients such as Claude Desktop and Cursor without code. Building the app itself with Cursor, Claude Code or Lovable? See Vibe Coding Web Search.

How were these prices checked?

Competitor prices come from each vendor's own pricing page, checked 2026-10-23 (Firecrawl and Jina Reader 2026-08-16) and linked above. Prices in this category change often, so check the vendor's page before you commit.

Try Serpex

Sign up for 200 free credits, no card required, and put Serpex through the evaluation steps above with your own queries. The docs cover the API, SDKs and MCP server, and pricing lists the rates by top-up, with credits that never expire.