Best Search API for AI Agents (2026): Tavily vs Exa vs Brave

Verdict: For most indie AI apps, start with Tavily or Exa, because both can hand your model cleaned page content in the same call. If the agent only needs ranked links, Brave Search API ($5 per 1,000 requests, own index) or Parallel (from $1 per 1,000) is cheaper. Use Serper or SerpApi only when the job specifically needs Google's results, and use a model's built-in search tool when speed of setup matters more than cost per query.

Prices come from each vendor's own pages, checked 13 Nov 2026. Advice is ours and marked as such.

TL;DR#

• Tavily: 1,000 free credits a month, then $0.008 per credit; a basic search is 1 credit.
• Exa: $10 of free credit each month, then $4 to $7 per 1,000 standard searches.
• Brave Search API: $5 per 1,000 requests, $5 of free credit each month.
• Cheapest per query: Serper (from $0.30 per 1,000) and Parallel (from $1 per 1,000).
• Built-in tools: $10 per 1,000 plus tokens on OpenAI and Anthropic; Gemini 3+ gets 5,000 free a month, then $14 per 1,000.
• Dead ends: Google Custom Search JSON API is closed to new customers; the Bing Search APIs are retired.

Disclosure: ZeroRaised is a launch site for bootstrapped products. We have no commercial relationship with any vendor in this post and earn nothing if you sign up. Facts come from the linked sources as checked on 13 Nov 2026.

What a search API does for an agent#

A search API tells your model which live pages match a query and what they say: grounding a chatbot, RAG over the live web, research agents, lead enrichment. Each returns one of three things, and that matters more than price:

1. Links and snippets. Cheap and fast; the model may still need to fetch pages.
2. Extracted content. Results plus cleaned page text or highlights.
3. An answer. A generated summary with citations.

Fetching known URLs, crawling or rendering JavaScript is a scraping job: see our web scraping API guide. A common pattern is search first, then scrape only the two or three URLs the model picks.

Comparison table#

"Not stated" means the vendor page we checked did not say, not that the answer is no.

The APIs one by one#

Tavily#

A search call can return snippets, cleaned page content (include_raw_content) and a generated answer (include_answer), up to 20 results (Tavily search reference). Basic, fast and ultra-fast cost 1 credit, advanced 2. Plans run from Project ($30 for 4,000 credits) to Growth ($500 for 100,000) (Tavily credits).

Our read: the easiest "search plus content" default. Bills grow on advanced depth and the Research endpoint (15 to 250 credits per pro request).

Exa#

Exa returns URLs plus optional full text, highlights and summaries, and says Lite mode serves content "directly from Exa's regularly refreshed search index" (Exa search reference). Per 1,000 requests: $4 for Lite/Instant, $7 for auto/fast, $12 to $15 for Deep Search, plus $1 per 1,000 pages of contents and $1 per 1,000 results above 10 (Exa pricing).

Our read: the best fit for "find pages like this" research and list building. Count the contents charge in your estimate.

Brave Search API#

Brave runs its own index. Search is $5 per 1,000 requests at up to 50 queries per second; Answers is $4 per 1,000 plus $5 per million tokens; both include $5 of free credit a month. Its FAQ says storing results or training models requires a custom plan, and "Full-funnel Zero Data Retention" is an Enterprise feature (Brave Search API).

The New York Times is suing Brave over its index and API; see our NYT vs Brave explainer. Our read: a solid independent index, but keep a second provider ready whichever you pick.

Serper and SerpApi#

Both return Google results as JSON. Serper sells prepaid packs: 50k credits at $1.00 per 1,000, down to $0.30 per 1,000 on the 12.5M pack, valid 6 months, with 2,500 free queries (Serper). SerpApi has 250 free searches a month, then $25 for 1,000, $75 for 5,000 or $150 for 15,000, counting only successful searches (SerpApi pricing).

Our read: use these when the job is "what does Google show" (SEO tools, rank checks, local results). You get links and snippets, so pair them with a scraper, and read each vendor's terms first.

Perplexity#

Perplexity's Search API is $5 per 1,000 requests, or $1 per 1,000 with Fast Search (Perplexity pricing). It returns "real-time, ranked web results as structured data" with title, URL, snippet and date, up to 20 results per request, and no generated answer (Perplexity search quickstart). Its Agent API bills model tokens plus tool fees, such as $0.0025 per web_search call. The pricing page did not list Sonar prices today, so we left them out.

Parallel and Linkup#

Parallel's Search API is "From $1 per 1k requests" with 10 results each, plus $5 free each month with a card on file and up to $80 at signup; Extract is $1 per 1,000 URLs (Parallel pricing). Linkup charges $0.005 to $0.006 per search, offers 4,000 free queries, and lists zero data retention as "available with the Enterprise plan" (Linkup pricing).

Jina and You.com#

Jina's s.jina.ai returns "the top five results with their URLs and contents". Each request costs tokens "starting from 10,000 tokens", new keys get 10M free tokens, and a key is required (Jina Reader). You.com charges $5 per 1,000 calls, gives $10 of free credit, and has a free MCP tier of up to 100 queries per day with no key (You.com pricing).

Built-in search from model providers#

• OpenAI: web search is $10 per 1,000 calls plus search content tokens at model rates; the preview tool on non-reasoning models is $25 per 1,000 with content tokens free (OpenAI pricing).
• Anthropic: $10 per 1,000 searches plus tokens; results count as input tokens on later turns, max_uses caps searches per turn, and failed searches are not billed (Claude web search tool).
• Google Gemini: Gemini 3 and newer share 5,000 free search requests a month, then $14 per 1,000, capped at 1,500 requests per day; Gemini 2.5 models are $35 per 1,000 grounded prompts after their free allowance (Gemini pricing).

Our read: built-in tools win on setup time and lose on cost and portability; results re-sent as input tokens on later turns can cost more than the searches.

Google Custom Search and Bing: do not start here#

Google says the Custom Search JSON API is closed to new customers and existing customers "have until January 1, 2027" to move (Google). Microsoft retired the Bing Search APIs on 11 August 2025 in favour of Grounding with Bing Search in Azure AI Agents (Microsoft Learn).

Decision guide by use case#

This is our editorial advice, based on the published prices and features above.

Our Model Context Protocol explainer covers how search MCP servers plug into a client.

Worked cost example: 10,000 searches a month#

Example, not a quote. An agent does 5 searches per task across 2,000 tasks a month: 10,000 searches. Our arithmetic from today's list prices, standard searches only, ignoring tax, your own model's tokens and content charges.

Tavily on advanced depth doubles to 20,000 credits. The spread from cheapest to dearest is more than tenfold, so choose on purpose.

Caching and query-budget tips#

1. Cache by normalized query plus a freshness window (an hour for news, a week for evergreen facts). Repeats then cost nothing.
2. Cap searches per task with max_uses on Anthropic's tool or a counter in your loop.
3. Ask for fewer results. Start with five and raise it only if quality drops.
4. Search wide, read narrow. Fetch content only for the pages the model picks.
5. Summarize before the next turn so old results do not ride along as input tokens. Our guide to reducing LLM costs covers this and prompt caching.
6. Put search behind one search(query, n) function so switching providers is a config change.

A note on benchmarks#

Several vendors publish search-quality benchmarks, and most comparison pages ranking for this query are vendor-written or dated, so we quote no benchmark numbers. Our read: check who ran a benchmark and on which queries, then run 50 of your own real agent queries through two or three APIs on their free tiers before you commit.

Frequently asked questions#

What is the best search API for an AI agent?#

For most indie developers, Tavily or Exa, because both return cleaned page content in the same call. Pick Brave for an independent index at $5 per 1,000 requests, Parallel for the lowest cost (from $1 per 1,000), and Serper or SerpApi when you need Google's results.

Which web search API has the best free tier?#

Exa resets $10 of credit every month, Tavily gives 1,000 credits a month, Brave and Parallel give $5 a month (Parallel needs a card), and Gemini 3+ includes 5,000 grounded searches a month. Serper and Linkup give one-off allowances of 2,500 and 4,000 queries.

Should I use OpenAI's or Claude's built-in web search instead of a search API?#

Use it for the fewest moving parts. Both charge $10 per 1,000 searches plus search content as tokens. A separate search API is usually cheaper, works with any model, and lets you cache results yourself.

Can I still use the Google Custom Search JSON API or Bing Search API?#

Not for a new project. Google says the Custom Search JSON API is closed to new customers, with existing customers given until 1 January 2027. Microsoft retired the Bing Search APIs on 11 August 2025.

How much does web search cost for an agent doing 10,000 searches a month?#

As an example at today's list prices: about $5 on Parallel after its monthly credit, $45 on Brave, $60 on Exa auto/fast, $72 on Tavily basic pay-as-you-go, $70 on Gemini 3 grounding, and $100 plus tokens on OpenAI's or Anthropic's built-in tool.

Launch what you build#

ZeroRaised is a launch directory for products built without institutional VC. When your agent is ready, run through our product launch checklist for bootstrapped makers, then submit your product for free. For more tool comparisons and guides on building without investors, join the ZeroRaised newsletter.

Sources#

• Tavily API credits and pricing, last checked 13 Nov 2026
• Tavily search endpoint reference, last checked 13 Nov 2026
• Exa pricing, last checked 13 Nov 2026
• Exa search reference, last checked 13 Nov 2026
• Brave Search API, last checked 13 Nov 2026
• Serper, last checked 13 Nov 2026
• SerpApi pricing, last checked 13 Nov 2026
• Perplexity API pricing, last checked 13 Nov 2026
• Perplexity Search API quickstart, last checked 13 Nov 2026
• Parallel pricing, last checked 13 Nov 2026
• Linkup pricing, last checked 13 Nov 2026
• Jina Reader and s.jina.ai, last checked 13 Nov 2026
• You.com API pricing, last checked 13 Nov 2026
• OpenAI API pricing, last checked 13 Nov 2026
• Anthropic web search tool, last checked 13 Nov 2026
• Gemini API pricing, last checked 13 Nov 2026
• Google Custom Search JSON API overview, last checked 13 Nov 2026
• Microsoft: Bing Search APIs retiring on August 11, 2025, last checked 13 Nov 2026