One agent task, twenty search calls: a cost model for agentic search pricing

CloudZero's Perplexity API pricing roundup, published 4 May 2026 against rates it dates to April, records two moves in opposite directions. Citation tokens, which were billed across the Sonar line in 2025, now apply only to Deep Research. At the same time a Pro Search mode carries a request fee the tracker puts at $14 to $22 per 1,000 queries, and an Agentic Research API adds roughly $0.005 per web search on top of the underlying OpenAI, Anthropic, Google or xAI model at that provider's own rate. I could not confirm the effective dates against Perplexity's own documentation, so treat the numbers as tracker-sourced rather than as a quote from the vendor.

The direction is the part that survives the hedge. The cheap unit got cheaper and a second, more expensive unit appeared beside it on the same price page, billed in a different currency. Perplexity Sonar is not unusual here. Through the first half of 2026 most serious vendors split their list the same way: a commodity search call priced per thousand requests, and an agentic mode priced per thousand requests at ten to fifty times the rate, or per run, or per compute unit.

Our May pricing post argued that vendor units do not line up and that you have to normalise before you compare. That is now harder, because the two SKUs you need to compare frequently belong to the same vendor. What follows is the arithmetic no published price page lets you compute: what an agent loop costs when one user task fans out into twenty-odd billable calls.

Four units, no common denominator

As of July 2026 there are three billing units in general circulation and a fourth arriving underneath them.

Tokens. Perplexity bills Sonar and Sonar Pro per million input and output tokens. Cost tracks context length, not call count, which makes it the one unit that gets worse as your prompts grow and better as your agent gets terser.

Requests. Exa publishes $7 per 1,000 searches for the base call returning up to ten results, $1 per 1,000 for results beyond ten, and $1 per 1,000 pages for contents. Serper.dev sits at the bottom of the same unit for raw Google SERPs. Brave announced an AI Grounding plan priced at $4 per thousand web searches plus $5 per million tokens, which is a hybrid: request-metered access with a token meter bolted on for the grounding step.

Runs and credits. Parallel AI prices its Task API per thousand completed runs on a ladder from $5 at the lite processor to $2,400 at ultra8x: a 480-fold spread inside one product on one unit, selected by a single parameter in the request body. Firecrawl uses credits, where a credit buys a fetch plus rendering plus markdown conversion.

Answers. On 1 July 2026 Cloudflare announced Pay Per Use, framed as an evolution of Pay Per Crawl, and pointed to Ceramic's pay-per-query model as an example: opted-in publishers get paid when their content appears in a search result rather than when a bot fetches the page. That moves the meter from the fetch to the answer. It is an announcement, not a settled market rate, but it is the first unit here that nobody's spreadsheet has a row for.

You cannot compare a $7 search, a $5 run, a credit and a token without a model that converts all four into cost per completed user task. So build one.

The model

Five variables. All of them are yours, not the vendor's, which is why no price page can compute this for you.

• T – tasks per day. A task is one thing a user asked for.
• S – search calls per task. How many times the agent's plan loop decides to query.
• F – fan-out. Query variants issued per plan step. Agents that rewrite one question into several phrasings multiply here.
• R – retry multiplier, 1 + retry rate. Empty result sets, timeouts, and reformulations on hard queries.
• P – pages fetched per search whose contents you actually pull.

Billable searches per task is S × F × R. Billable fetches per task is S × F × R × P.

Take a research assistant doing 1,000 tasks a day, with six plan steps, three query variants per step, a 25% retry rate, and three pages pulled per search. That is 22.5 billable searches and 67.5 billable fetches per task, or 675,000 searches and about 2.03 million fetches a month.

At Exa's published rates that is $4,725 for the searches and $2,025 for the contents, so roughly $6,750 a month, or $0.225 per task. The search line is 70% of it, which is not where most teams look first.

Now run the same loop against a per-request agentic tier. If Perplexity's Pro Search request fee fires once per search call rather than once per task, 675,000 requests at $14 to $22 per thousand is $9,450 to $14,850 a month before a single token. Same workload, same twenty-odd calls, roughly double to triple the bill because of which SKU the loop happened to hit.

The comparison inverts when you multiply

Exa's published Deep Search rate is $12 to $15 per 1,000 requests, nominally 2× the base search. If one Deep Search call replaces the entire twenty-two-call loop, 30,000 tasks a month costs about $405. That is roughly 16× cheaper than hand-rolling the loop out of $7 searches, using the same vendor's own price page.

The expensive SKU is priced per task. The cheap SKU is priced per call. An agent makes many calls per task. So the "premium" agentic tier can be the cheaper line item, and the "commodity" search tier can be the one that runs away from you, and the direction depends entirely on your S × F × R, which appears nowhere in either price.

This also explains where bills actually blow up. Three failure modes account for most of the surprises:

1. The agentic tier fires inside the loop instead of replacing it. You pay task-level pricing at call-level frequency. This is the difference between $405 and $12,000 a month in the example above.
2. Retries are billed and invisible. A 25% retry rate is a 25% surcharge on every line. Some vendors are explicit about the treatment: Parallel's documentation states that failed runs are not billed and you pay only for successfully completed runs. Most search APIs bill a zero-result response the same as a good one, and a query that returns nothing is exactly the query your agent will reformulate and re-fire.
3. Fetch cost stacks on search cost. Exa's $1 per 1,000 pages looks like a rounding error next to $7 per 1,000 searches, until P = 3 makes it 30% of the bill. Anything that pulls full page contents, rather than snippets, is a second meter running in parallel with the first.

Cloudflare added a number on 1 July 2026 that sits directly on top of this: more than 50% of crawl traffic from good bots, it said, goes to re-fetching pages that have not changed. On any per-fetch plan that is billable volume buying you nothing. A conditional-fetch cache keyed on content hash removes most of that volume and takes about a day to build.

Free credits are not a price signal

On 16 June 2026 Firecrawl shipped Keyless: search and scrape with no account and no API key, and 1,000 free credits a month granted automatically. It is a good developer experience decision and a bad input to a cost model.

A keyless or generous free tier tells you the vendor's marginal cost per fetch is low enough to give some away as acquisition spend. It tells you nothing about your production line item, because the free allowance is denominated in the cheap unit and your bill will be denominated in whatever unit your loop actually hits. A thousand free credits covers roughly fifteen tasks in the model above. Prototype inside a free tier and scale the number linearly and you are extrapolating from the wrong unit.

The same caution applies to comparing entry tiers across Tavily, Exa and the rest. The comparison that decides anything happens at your volume with your fan-out.

The line item nobody has modelled

If Pay Per Use style arrangements spread, your extraction vendor starts paying publishers when their content appears in an answer, and that cost passes through to you. The unit will not be the one you budgeted. You will have modelled fetches, or searches, or runs, and the pass-through will be metered on answers, which correlates with how often a source is useful rather than how often it is retrieved.

A high-fan-out agent that fetches broadly and cites narrowly gets cheaper under answer-metering and more expensive under fetch-metering. An agent that fetches precisely and cites everything gets the opposite. We covered the payment rails themselves in the pay-per-crawl post; the cost-model consequence is that citation rate becomes a budget variable, and today nobody instruments it.

Five questions before you sign

Ask these in the procurement call. None of them are answerable from a public price page.

1. What triggers the agentic tier, and is it opt-in? If a router can escalate a call to the expensive SKU on its own judgement, your unit cost is set by the vendor's classifier, not by you.
2. Are retries and zero-result responses billed? Get it in writing. It is a 10 to 30% swing.
3. What is the fetch cost inside a search? Ask whether contents, snippets, summaries and reranking are inside the search price or metered separately. Two vendors now list the same extra line: Exa's agent pricing and Perplexity's Agentic Research API each add about $0.005 per web search on top of everything else.
4. What is the overage rate, and does it step or cliff? Credit packs that expire monthly and overage priced above list are how a predictable subscription becomes an unpredictable one.
5. What is the rate at my committed volume? Above roughly $2,000 a month the published price is a starting point, and most vendors in this category quote against an annual commitment.

What this means if you build on web data

Model S × F × R before you compare vendors. The multiplier between one user task and one billable call is the entire decision. Two teams reading the same price page will get bills that differ by an order of magnitude, and neither price page is wrong.

Instrument searches per completed task in production, not in the prototype. It is the one number that converts every vendor's unit into a comparable figure, and a prototype loop is always shorter than a production loop that has to survive bad queries, empty result sets and reformulations.

Treat the agentic tier as a candidate for the cheap option, not the expensive one. Per-task pricing beats per-call pricing whenever your loop is long. Price both shapes against your own multiplier before deciding; our comparison of AI search APIs covers which vendors sell which.

Add a cache and a citation counter now. Cloudflare's re-fetch figure says half of crawl volume is wasted, and answer-metering, if it arrives, will bill on citations you currently do not count. Both are cheap to add today and expensive to retrofit once they are on the invoice. If that pushes you toward owning more of the stack, the build-versus-buy arithmetic has moved with it.