Cost per completed task

Why the per-call price misleads

Two providers at the same per-call price can differ several-fold per task. One returns a set the agent can act on; the other returns a set the agent has to widen, re-query and read in full before it can answer. The second makes more calls, consumes more context and fails more often — and its pricing page looks identical.

The reverse also holds. A call that costs more but replaces three — one ground call instead of a search, two fetches and a corroboration loop — can be the cheaper task.

The formula

Run a fixed set of tasks. Record every call and every model token. Grade each task as completed or not against criteria written before the run. Then divide.

provider spend  = sum over tasks of (calls × price per call)
model spend     = sum over tasks of (input tokens × input price + output tokens × output price)
completed       = tasks graded acceptable against criteria set before the run

cost per completed task = (provider spend + model spend) / completed

A worked example, on our own prices

The research-agent recipe on this site makes 1 authorities + 1 dossier + 2 related + 1 path per objective — 10 credits. At 10,000 objectives a month that is 100,000 credits, which is cheapest on Build: $79 a month including overage, or $7.90 per 1,000 objectives in provider spend.

That is one line of three. Model spend depends on your model and on how much of each result set it reads, which is why the response shape matters as much as the price: a budget-packed evidence set is fewer input tokens than ten page bodies. Review effort is the third line, and it is the one a weak evidence set inflates most.

The call pattern is illustrative — a starting point, not a measurement. Yours is whatever your agent actually does, which is what the run above records. The pricing page costs every published pattern at three volumes.

Retries, thin results and the cost of not stopping

A call that runs and fails costs 1 credit here, whatever it was; a refused call costs nothing. The larger cost is behavioural: an agent handed a thin result set searches again, and again, because nothing told it the evidence does not exist in the index.

An explicit status is cheaper than a loop. When ground returns insufficient with next actions, the agent can stop and abstain on the first call instead of spending five more discovering the same thing.

Keep three lines separate

Provider spend, model spend and review effort move independently. A retrieval change that lowers provider spend and raises review effort has not saved anything, and a single blended figure would hide that. Report all three per completed task, and compare them before and after any change.