Last updated: 2026-06-22

Agent cost explained

The cost of an agentic run is the combined price of the sequence of large language model calls an agent makes to complete one task. An agent alternates between reasoning, calling a tool, and incorporating the result, with each model invocation counting as a step. Because every step is billed for the tokens it consumes, the run cost is the per-step cost multiplied by the number of steps — typically several times the cost of a single direct call.

Why multi-step runs cost more

A one-shot request bills for one prompt and one completion. An agent loop bills for both at every step, and the input side grows as the run proceeds: each step re-sends the accumulated context — the original instructions, prior reasoning, and every tool result so far — so the model can act with full history. A task that a single call would resolve for a fraction of a cent can cost many times that once the same context is paid for across a dozen steps.

The formula

With $n_{in}$ input and $n_{out}$ output tokens per step at per-million prices $p_{in}$ and $p_{out}$, the cost of one step is

and a run of $N$ steps costs

This treats every step as consuming the same token count, which keeps the arithmetic simple at the cost of ignoring the growth of context across the run.

Worked example

Consider an N = 8 step run where each step sends 4000 input tokens and generates 600 output tokens, priced at $3 and $15 per million. The per-step cost is

so the run costs C = 0.021 \times 8 = \0.168$. The large input figure reflects context being re-sent each step; here the 4000 input tokens at the lower rate still contribute more than half of each step's cost, and the eight-fold repetition is what separates an agent run from a single call.

Notes and variations

The uniform-step assumption understates real cost when context accumulates, since later steps carry more input than earlier ones. For a precise figure, sum the steps individually rather than multiplying an average. Prompt caching is especially effective here, because the stable portion of the context — the system prompt and early history — repeats at every step and can be billed at a reduced rate after the first call.

Application

Agent cost extends the per-call model to workflows that loop. The single-step price it builds on is the subject of the Token Cost Calculator, and scaling a run to a daily or monthly fleet of tasks is handled by the LLM Monthly Cost Calculator. Because re-sent context dominates the input bill, Prompt Caching Savings Calculator quantifies the discount caching can recover.