A cost of zero carries no information. If calling an endpoint is free, an agent (or whatever loop is driving it) has no reason encoded in the system itself to call it once instead of fifty times. A cost of $0.01, even one that small, changes that completely - every call now has a consequence that shows up somewhere, which means the system has a signal it didn't have before.
Consider a retry loop that's slightly broken - an agent re-drafting the same email because a downstream check keeps failing for an unrelated reason, or re-polling an endpoint far more often than the task needs. Against a free API, that loop is invisible until something else catches it: a timeout, a rate limit, a human noticing the logs look busy. Against a priced one, it's visible immediately, as a cost that accumulates in direct proportion to how often the loop actually fires. A $0.02 draft call made by mistake forty times costs $0.80 - small, but a real, attributable number instead of nothing.
It's not just free APIs that lose this signal. A flat monthly plan with an unused quota has it too: once the quota's paid for, the marginal cost of one more call inside it is also zero, for exactly the same reason a free API's marginal cost is zero. Subscription pricing and free pricing fail the same way here - both decouple the cost of an individual action from the decision to take it. Pay-per-call pricing is the one shape that keeps marginal cost real on every single call, whether it's the agent's first of the day or its ten-thousandth.
DimeBox never charges for a failed call - a response under 400 is the only thing that settles payment. That's a deliberate complement to charging in cents, not a contradiction of it: the cost signal is specifically about successful, billable work, not about punishing an agent for a validation error it's going to fix and retry anyway. An agent iterating toward a correct request shape - missing a required field, trying a reserved slug, getting the content type wrong - pays nothing while it's wrong and pays once it's right. That keeps the pricing signal aimed at genuine overuse, not at the ordinary trial-and-error of a machine caller finding the right request shape.
Retries raise a separate question worth distinguishing from the behavioral one: a network-level retry of a call that already succeeded, where the response just got lost, shouldn't bill twice for one action. That's what idempotency keys are for - a different mechanism, aimed at making retries safe rather than at making overuse visible.
A human using Gmail doesn't evaluate the marginal cost of hitting send - it's a reflex, bundled into whatever they're already paying for their email account. An agent sending through a priced endpoint can't treat it as a reflex, because every send is a line item. That's arguably a better default for autonomous outbound communication than reflexive, costless sending would be: free outbound email from an agent is a standing invitation to spam, and attaching even a small, real price to every send forces whatever's driving the agent to treat "should I send this" as a decision with a cost attached, not something it does by default.
Think about what changes for the operator watching from outside, not just the agent itself. A free endpoint produces no record worth looking at - every call looks identical, costing nothing, whether it did something useful or not. A priced endpoint turns every call into a line item with a dollar amount attached, which means a spike in calls shows up as a spike in spend before it shows up as anything else. That's a cheaper, earlier signal than waiting for a rate limit to trip or a server to slow down.
This is also why pricing the action rather than pricing access matters specifically for autonomous loops. A human operator glancing at a dashboard once a day can catch a runaway script eventually. An agent running unattended for hours doesn't have that glance happening in real time - the per-call price is effectively doing the watching, continuously, by making every unnecessary call cost something the moment it happens rather than at whatever point a human next checks in.
None of this depends on the price being large. A cent or two is not meant to be a meaningful deterrent on its own - it's meant to be nonzero, consistently, on every call, so that cost and usage never fully decouple the way they do under both free and flat-rate pricing. The behavior changes not because the price hurts, but because the price exists at all, on every single call, every single time.