Search API for AI Agents Priced: $300 to $2,400 a Month at 10,000 Queries a Day

An agent that runs 10,000 web searches a day buys about 300,000 queries a month. On the published rate cards, that bill runs from $300 on Perplexity's Fast Search to $2,400 on Tavily's pay-as-you-go basic search. Choosing a search API for AI agents is an eightfold price decision before result quality even enters the room. Price is also the easier half of the decision.

Short answer

For 10,000 agent queries a day, the cheapest published option is Perplexity's Search API with Fast Search at $1 per 1,000 requests, about $300 a month. Exa costs $2,100 but includes page text. Brave sits near $1,495. Tavily pay-as-you-go is $2,400 for basic search and $4,800 for advanced.

Search API for AI agents: the four rate cards side by side

Every research agent eventually needs a web search API. The model cannot browse on its own, and scraping a consumer search engine breaks its terms. So the search call becomes a metered line item that scales with every task your agent runs.

The four vendors below all sell to that use case. They do not price it the same way. One bills in credits, one in requests that can hold several queries, and two in flat requests per 1,000. Normalising them to a single unit is the first job.

Every row comes from the vendor's own pricing page, changelog or docs. None comes from a comparison site. Enterprise tiers for all four are priced on request and excluded.

Exa: one price, contents included

Exa simplified its pricing on 3 March 2026. A search with contents now costs $7 per 1,000 requests and returns 10 results with text and highlights attached. Results beyond 10 cost $1 per 1,000. Summaries cost another $1 per 1,000 pages.

On 19 October 2026 Exa added a lite search type at $4 per 1,000 requests. Lite never fetches a page live, so it trades freshness for price. Deep search sits at $12 per 1,000 requests. The free tier grants $10 of credit a month, which is about 1,400 default searches.

Tavily: credits with a plan ladder

Tavily bills in credits. A basic search costs 1 credit and an advanced search costs 2. Pay-as-you-go is $0.008 per credit, so 1,000 basic searches cost $8. The monthly plans in Tavily's docs bring the credit price down to $0.005 on the $500 Growth plan, which buys 100,000 credits.

That ladder is the catch at agent volume. The largest published plan covers a third of a 300,000-query month. Anything beyond it runs at pay-as-you-go rates or needs an enterprise quote. Tavily's own pricing page lists only the free, pay-as-you-go, $30 Project and enterprise options, so the cheaper middle tiers appear in the docs rather than on the page most buyers read first.

Perplexity: per request, with batching

Perplexity's Search API returns raw ranked results at $5 per 1,000 requests, or $1 per 1,000 with Fast Search switched on. The billing note is the interesting part. A successful request containing up to five queries counts as one billing unit, and there are no token charges on top.

Failed, invalid and rate-limited requests are not billed either. That is a small detail with real money behind it, because agents retry.

Brave: one flat price per family

Brave relaunched its API pricing on 12 February 2026. Every search type, from web results to the LLM Context endpoint, costs $5 per 1,000 requests. Each plan includes $5 of free credit a month, which Brave says is conditional on attributing the API on your site.

Brave also sells an Answers plan at $4 per 1,000 requests plus $5 per million input and output tokens. It is a different purchase from raw search.

Perplexity Sonar API pricing changed in September 2026

If you searched for Perplexity Sonar API pricing, the tables you probably found are out of date. Comparison sites still quote Sonar request fees split by search context size. The current Perplexity pricing page, read on 27 October 2026, carries no such table.

The reason is structural. Perplexity's migration guide states that Sonar Chat Completions support ended on 27 September 2026. Synchronous and streaming requests still work, because Perplexity reformulates them as Agent API requests. Asynchronous Sonar requests are no longer supported.

Under the Agent API, Sonar is a model like any other. The models page lists perplexity/sonar at $0.25 per million input tokens and $2.50 per million output tokens. Web search becomes a tool, billed at $2.50 per 1,000 invocations, or $1 with Fast Search, on top of those tokens.

My view is that this change is good for buyers. A bundled request fee hid how much of the price was search and how much was model. Splitting them lets you see that the search component of a Sonar-style answer now costs half of Perplexity's standalone Search API. It also means any budget built on the old fee tables needs rebuilding, and that pattern is tracked across vendors in the calendar of API price changes and sunsets.

What a 10,000-query-a-day agent costs on each search API

The scenario is deliberately plain. One agent, 10,000 search calls a day, 30 days, so 300,000 queries a month. Each query wants the default result count. No enterprise discount, no negotiated rate, and free monthly credits are netted off where they apply.

Computed by Zan Digital from the published rates in the first table. Rows are not equivalent products. Exa includes page text, the others mostly return snippets, so read the next section before treating the cheapest row as the winner.

The batching discount is the largest single lever

Perplexity's per-request billing rewards an agent that plans its queries before it fires them. A research agent that decomposes a question into five sub-queries can send all five in one request. At Fast Search rates, the month drops from $300 to $60.

The limit is architectural. Batching only works when your agent knows several queries at once. An agent that searches, reads, then decides the next search cannot batch, because each query depends on the last. If your loop is sequential, plan on the unbatched row. The broader problem of agents billed on meters you cannot predict in advance is covered in the breakdown of agent task costs across vendor meters.

Advanced modes double the bill quietly

Tavily's advanced search costs 2 credits instead of 1, which moves the month from $2,400 to $4,800. Exa's deep search at $12 per 1,000 takes the same workload to $3,600.

The risk is that the mode is a single parameter in your agent code. A developer who flips search_depth to advanced during a quality fix has doubled a line item without a purchase order. I would put the search mode in configuration that someone with budget authority reviews, not in a prompt or a default argument.

Content extraction is where Tavily vs Exa pricing diverges

A search result is a title, a URL and a snippet. Most agents need more than that. They need the text of the page to quote, summarise or check a claim. How each vendor prices that second step changes the ranking in the table above.

Exa folds it into the base price. Since March 2026 the $7 search returns text and highlights for the first 10 results. If you call Exa's separate contents endpoint, it costs $1 per 1,000 pages per content type, so text and highlights on one page count as two.

Tavily charges extraction as its own credit line. A basic extract costs 1 credit for every 5 successful URL extractions, and failed extractions are not charged. An agent that searches and then reads the top 5 pages spends 2 credits per task. On pay-as-you-go that is $16 per 1,000 tasks, against Exa's $7 for 10 pages of text.

Brave's answer is an endpoint it calls LLM Context, priced the same $5 as web search. Perplexity's Search API is described as raw ranked results with advanced filtering. I could not verify from the docs alone how much page text either returns per result, so test both against your own prompts before assuming they replace a reader.

If your agent reads pages, compare cost per read task, not cost per search. On that measure Exa moves from the expensive half of the table to near the cheapest content-inclusive option among the four.

Rate limits matter for bursts, not for daily volume

Ten thousand queries a day averages about 0.12 queries per second. Every vendor here clears that by a wide margin. At its default limit of 10 queries per second, Exa could serve 864,000 queries a day, 86 times the scenario.

The constraint is the burst. A research agent that fans out 30 parallel searches in one second exceeds Exa's default of 10 per second and gets 429 errors. Exa raises the limit to 25 per second for teams that add $1,000 of credit inside 30 days. On Tavily's per-minute limit of 1,000 requests on a production key, the same burst uses 3% of a minute's allowance, assuming the window is a simple count, which the docs do not detail.

Perplexity's Search API allows 50 query units per second on every usage tier. Note that billing and limits count differently: a five-query request is one billing unit but five units against the rate limit. Its Agent API is tighter for new accounts, starting at 1 request per second on Tier 0 and reaching 3 after $50 of cumulative purchases, per its usage tiers page.

Tavily adds one more condition worth knowing before launch. Its rate limits page says production keys require a paid plan or pay-as-you-go. A development key is capped at 100 requests per minute. Why vendors publish limits this unevenly is examined in the case for a rate limit transparency standard.

Data retention terms run in two directions

Search is an unusual data flow. Your agent sends the vendor its queries, and the vendor sends back content it has licensed or crawled. Both directions carry terms. Most comparisons check neither.

What the vendor may do with your queries

Your queries are not neutral. An agent working on a customer's acquisition target, a legal matter or an unreleased product sends that context out in plain text. Exa's privacy policy, updated 29 June 2026, says query data is used to improve its products, including by training and fine-tuning models. Zero data retention exists, but only on enterprise plans.

Tavily's terms of service, section 6.5, allow Tavily and its third-party AI providers to retain customer input for training. Its privacy policy adds that query data may go to third-party index providers when Tavily's own index cannot answer. Perplexity's privacy page states it does not train on customer content under its API terms, though it says zero retention is not the default. How the model providers behind your agent compare on the same question is ranked in the ranking of AI vendor data retention terms.

What you may do with their results

Brave's terms of use, updated 1 September 2026, bar customers from storing or caching search results beyond the transient storage needed to run an application. They also bar using results to train, fine-tune or benchmark AI models. That rules out the common pattern of logging every search result into an evaluation set.

I did not read the equivalent clauses for Exa, Tavily and Perplexity closely enough to state them, so those cells say so. If your agent keeps a cache of retrieved pages to cut costs, read this clause on every vendor before you build the cache. A cache that saves 40% of your search bill is worthless if the contract forbids it.

Where this comparison is weakest

A rate card comparison has two blind spots, and this post has both.

The first is quality. Price per 1,000 queries says nothing about whether the results answer the question. A cheaper API that forces your agent into two extra searches per task is more expensive per task. Zan Digital has not run a retrieval benchmark across these four, and no figure here measures relevance, freshness or recall. The failure that matters most, an agent confidently citing a wrong page, is the subject of the audit of hallucinations in deep research tools.

The second is decay. Three of the four vendors changed pricing structure inside 2026. Exa rebundled in March and added lite in October. Brave dropped its free tier in February. Perplexity retired Sonar Chat Completions in September. Every figure here was read on 27 October 2026 and should be treated as dated from that day.

There are smaller gaps too. The Sonar-style cost depends on token counts that vary by prompt, so this post gives the tool fee and leaves tokens open. Brave's 50 queries per second figure comes from its pricing page rather than the terms. Tavily's cheaper plans appear in its docs but not on its pricing page, which leaves some doubt about which list a sales team will honour.

Which search API to choose for which agent

There is no single winner. Here is how I would choose, by the shape of the agent.

An agent that reads pages on every task. Start with Exa. The $7 price with text for 10 results is cheaper per read than search plus a separate extract elsewhere. Accept that your queries feed its models unless you buy enterprise.

A high-volume agent that only needs links and snippets. Start with Perplexity's Search API on Fast mode, and batch wherever your planner allows it. At $1 per 1,000 requests, the search line stops being worth optimising, and the no-training default is a bonus.

An agent handling sensitive queries. Shortlist Brave and Perplexity, both of which say they do not train on your queries. Get the zero retention commitment in writing before production, because neither publishes it as a self-serve switch with a stated retention period.

A small agent under 100,000 searches a month. Tavily's plans become competitive here. The $500 Growth plan works out to $5 per 1,000 basic searches, the same as Brave. Stale results are a separate failure from cost, and the analysis of stale agent answers covers how to test for them.

Frequently asked questions

Where to start this week

Pull one week of your agent's search logs and count three things: queries per task, how many tasks read page text, and the peak queries in any single second. Those three numbers decide which row of the cost table is really yours, and most teams have never measured them.

Then send the same 50 real queries from those logs through two of these APIs and compare the answers your agent produces, not the raw results. Price the winner per completed task. If the switching cost worries you, route search through one internal function so the vendor becomes a configuration line, the same pattern behind the comparison of AI gateway fees and failover.

Related reading