Comparison, updated 30 August 2026

Cheapest web search API for AI agents in 2026, priced per 1,000

Short answer: on published list prices, Serper is the cheapest per request, falling to $0.30 per 1,000 on its largest pack. Neurobird Search is $2 per 1,000 and is the cheapest that returns the extracted page content in the same call. Brave Search API and Google Custom Search are $5, Exa about $7, Tavily $8 basic and $16 advanced, and the Anthropic web search tool $10 plus token costs. If your agent only needs links, buy the cheapest SERP API. If it needs to read the pages, count that second fetch before you compare unit prices, because it is usually where the money actually goes.

Every figure below was read at the vendor's own pricing page and is linked in place. Prices marked August 2026 were read that month; Firecrawl and Linkup were read on 30 August 2026. This is a dated snapshot and vendors change prices, so follow the link before you budget against it. Neurobird Search is our own product, and this article says plainly where the other options beat it.

The table

Published price per 1,000 requests

ProviderPer 1,000Returns page contentFree tier
Serper$0.30 at 12.5M credits, $1.00 at 50kNo, links and snippets2,500 credits
Neurobird Search$2.00, $1.00 above 1M credits a monthYes, included1,000 a month, no card
Brave Search API$5.00NoLimited free plan
Google Custom Search$5.00, discontinued 1 Jan 2027No100 queries a day
Exaabout $7.00, contents billed separatelySeparate chargeTrial credits
Tavily$8.00 basic, $16.00 advancedPartial, snippets1,000 a month
Anthropic web search$10.00 plus token costsYes, in-modelNo
Linkup$1.00 to $6.00 depending on depthFetch endpoint4,000 queries

Firecrawl is deliberately absent from the per 1,000 column. Its published plan is $83 a month for 100,000 credits, which works out at about $0.83 per 1,000, but those are scraping credits rather than search requests, so putting it in the same column would compare two different things. It is the right tool when you already know which URLs you want.

The part the price table hides

Cost per request is not cost per answer

A SERP API returns ten links and a one line snippet each. That is genuinely cheap, and if your agent only needs to know which pages exist, it is the correct purchase. But an agent that has to answer a question then makes a second round of requests to fetch those pages, and that round is where the real cost sits: it adds seconds of latency, it fails on the sites that block automated readers, and it spends tokens on navigation and cookie banners before it reaches the paragraph that mattered.

So the honest unit is cost per useful answer. Multiply the unit price by the number of calls your agent actually makes for one question, and add what the fetch round costs you in failures and tokens. Sometimes the $0.30 option still wins. Often the thing that wins is whichever API removes the second round.

Picking one

Which to choose, by what is actually constraining you

Volume is the constraint, links are enough

Buy Serper. At $0.30 per 1,000 on the largest pack nothing else is close, and if you only need to know what exists, paying more buys you nothing.

A person is waiting on the response

Buy Tavily or Brave. Both are faster than we are. Our median is measured in seconds because a call opens and reads real pages, and that is the wrong trade when a human is watching a spinner.

You need semantic or neural retrieval

Buy Exa. It runs its own index built for meaning rather than keywords, which is a genuinely different capability, not a cheaper version of one.

Cost per useful call is the constraint

Try Neurobird Search. $2 per 1,000 with the page content in the same response, and 1,000 free credits a month to check the claim before paying.

Disclosure and method

How this was put together

Neurobird Search is our product, so treat our entry with the scepticism you would apply to any vendor writing its own comparison. What we can offer instead of a promise is a method you can repeat: every price here comes from the vendor's own published pricing page rather than a secondary source, each is linked so you can check it, and each carries the month it was read. Where a competitor is better at something we are not, the section above says so by name.

We also publish our own limits rather than burying them: there is no index of our own, so coverage and freshness are inherited from upstream engines, median latency is about 15 seconds basic and 28 seconds advanced, and there is no SLA and no SOC 2. Those are on our about page in the same words.