Web search API pricing per 1,000 searches

Web search API pricing in October 2026 differs up to fifteen times between vendors for the same 1,000 searches, and the search built into the big model APIs adds a token bill on top. This post compares Perplexity, Brave, Tavily, Exa, Cloudflare, OpenAI, Anthropic and Gemini, with each price linked to the vendor's own page and read on 2026-10-13.

Two things changed this month. Perplexity moved its Agent API presets to Fast Search at $1.00 per 1,000 (Perplexity changelog, read 2026-10-13), and Cloudflare launched a Web Search API inside AI Gateway with Exa, Linkup and Ceramic as partners (Cloudflare blog, read 2026-10-13). This post puts the prices in one table, explains why "one search" means different things per vendor, and shows how an agent reads the price before each call.

Web search API pricing in one table

The table lists the vendor's pay as you go price for 1,000 searches, the unit it bills and what one search returns. Free tiers and plans are left out unless they change the first 1,000 calls.

Tavily gives 1,000 free credits in each calendar month, so the first 1,000 basic searches cost nothing (Tavily credits, read 2026-10-13). Exa bills page text apart from the search, at $1 per 1,000 pages per content type (Exa pricing, read 2026-10-13). Cloudflare publishes no number of its own: it charges the partner's API price from your AI Gateway credit balance.

What does the search built into model APIs cost?

The model APIs charge per search and then bill the search results as tokens. So the per 1,000 price is a floor, not the bill.

The token line is the one people miss. Anthropic counts the results of every search as input tokens in that turn and in later turns of the conversation. A search that pulls 5,000 tokens of results can cost more in tokens than the search fee itself, at the model's input rate.

Why does web search API pricing use different units?

A search is a different unit per vendor, so compare what you get per call, not only the price. There are three kinds of answer.

1. Results only. Titles, links and snippets. Perplexity Search API and Brave Search return this, and your agent still has to read the pages.
2. Results with page text. Tavily returns content with each result. Exa returns results, and page text is a separate line on the bill.
3. An answer. The model API tools search, read and write an answer in one step, and you pay for the tokens that took.

If your agent needs the full text of 5 pages per question, a results-only API plus a page reader can cost less than a search that bundles text, or more. Count both calls before you pick. Reading a page as markdown is its own job, covered on the URL to markdown API page.

How do you pay per search without a subscription for each vendor?

Call the search providers through one prepaid balance and read the price before each call. That is what looot does for agents: the job `web.search` has several providers behind it, each with its own live price. The table below reads the cheapest price per call from the looot catalog when this page loads. Multiply by 1,000 to compare with the tables above.

Each provider has a page with its jobs and how to connect it to an agent: Exa, Tavily and Linkup. The job itself, with inputs and output, is on the web search API page. If you need Google's own results as rows rather than a web search, that is a SERP API, and its prices are in SERP API pricing.

This prompt compares the live prices for your own workload without spending anything:

What this post cannot tell you

• Search quality. We compare prices and units, not how good the results are. Run your own questions through two or three providers.
• Your token bill. For the model API tools, the search fee is a floor. Tokens depend on the model and how many results it reads.
• Volume discounts. Several vendors sell plans and enterprise prices. We list only the pay as you go price each vendor publishes.
• Cloudflare's numbers. Cloudflare charges each partner's list price and publishes no table of its own in the launch post.
• Scheduled searches. Running the same search on a timer needs scheduled runs, not available yet in looot.

Last checked 2026-10-13. Vendor prices were read that day from the pages linked in each row. looot prices in the live table are read from the catalog when the page loads. We will re-read every vendor price within 92 days.