Estimate what an AI agent costs per run and per month once you factor in every step of its loop. Adjust steps, tokens per step, and model — the estimate updates live.
Each agent run is made up of multiple steps, and each step consumes its own input and output tokens. Tokens per run are steps times input tokens per step (for input) and steps times output tokens per step (for output). Cost per run prices those tokens at the model's per-token input and output rates, and monthly cost multiplies that by your runs per month. Step count is the hidden multiplier here: a 6-step agent effectively sends 6x the token volume of a single non-agent LLM call at the same per-call token size, which is the number one reason agent budgets blow past expectations.
It doesn't account for prompt caching discounts on repeated context, retries on failed tool calls, or agents that dynamically vary their step count based on task difficulty. Treat this as a directional estimate for budgeting an agent workload, not an exact invoice.
Each agent run consists of multiple steps (LLM or tool calls). Tokens per run are steps times input and output tokens per step. Cost per run is those tokens priced at the model's per-token input and output rates, then multiplied by runs per month.
An agent that takes 6 steps to complete a task sends roughly 6x the tokens of a single-shot LLM call with the same per-step token size, because each step re-sends context and adds new reasoning or tool output. Step count is the single biggest cost multiplier in agent systems.
Trim the context window sent at each step, cache repeated context, use a cheaper model for simple sub-steps and reserve the strongest model for final reasoning, and cap the maximum number of steps per run to catch runaway loops.