AI Agent Run Cost Calculator
An agent does not make one model call: it plans, calls tools and reads the results over several steps, and every call is billed. Enter the average tokens per call, how many calls each step makes and how many steps a run takes, plus optional tool or MCP overhead and cached input, and see the cost per run, per day, per month and per year. The result is an estimate: the prices and token counts are yours, and provider extras are not included.
Result
Estimate
Enter the tokens, calls, steps, runs and prices to estimate the cost.
- Model calls per run
- Input tokens per run
- of which cached
- Tool overhead tokens per run
- Output tokens per run
- Total tokens per run
- Cost per run
- Cost per day
- Cost per month
- Cost per year (365 days)
Where the cost goes (per run)
| Part | Cost | Share |
|---|
Exact arithmetic on your numbers, but an estimate of a real bill: token counts vary from run to run, and provider extras such as tool or search fees, retries and taxes are not included.
Examples
Worked examples at example prices of 3 USD input and 15 USD output per million tokens, 100 runs a day, 30 days a month:
| Agent | Calls per run | Tokens per run | Cost per run | Cost per month |
|---|---|---|---|---|
| Simple assistant | 3 | 7,200 | 0.036 USD | 108.00 USD |
| Research agent with tools | 16 | 116,000 | 0.444 USD | 1,332.00 USD |
| Coding agent with caching | 20 | 420,000 | 1.07 USD | 3,204.00 USD |
Formula & Methodology
The arithmetic is exact for the numbers you enter:
calls per run = steps × calls per stepinput tokens per run = calls per run × input tokens per call; overhead = steps × overhead per stepoutput tokens per run = calls per run × output tokens per callcost per run = uncached input × input price + cached input × cached price + overhead × input price + output × output price (each ÷ 1,000,000)per day = cost per run × runs per day; per month = per day × days per month; per year = per day × 365Good to know
- Agents re-send their growing history on every call, so input tokens per call rise during a run. Enter the average over a whole run, or measure a real run and use its usage figures.
- Tool or MCP overhead is added once per step as ordinary input. Many clients send tool definitions with every model call, in which case multiply the overhead by the calls per step.
- Cached tokens are billed at the cached price you enter. Cache-write fees, storage fees and cache lifetimes depend on the provider and are not modelled: see the prompt cache calculator.
- Not included: tool, search and code-execution fees, retries, failed runs, batch discounts, taxes and free tiers.
- Costs vary from run to run. Use the estimate for planning and your provider’s usage reports for the real figure.
How to Use
- Enter the average input and output tokens of one model call.
- Enter the calls per step and the steps per run, plus any tool overhead or cached tokens.
- Enter the runs per day and your prices per million tokens, or prefill them from a preset.
- Read the cost per run, day, month and year and the breakdown.
Frequently Asked Questions
How do I estimate what an AI agent costs?
Count the model calls in a run (steps × calls per step), multiply by the average input and output tokens per call, price them per million tokens and add any tool overhead. Then multiply by the runs per day and days per month. This calculator does all of that.
Why is an agent run more expensive than one chat reply?
It makes many model calls, and each call re-sends the instructions, tool definitions and the growing history as input. Costs therefore grow faster than the number of steps.
How do tool and MCP definitions affect the cost?
They are extra input tokens sent to the model. This tool adds the overhead you enter once per step; if your client sends the definitions with every call, multiply it by the calls per step. The MCP overhead calculator estimates the definitions themselves.
Can prompt caching lower an agent’s cost?
Often, because much of each call’s input repeats. Enter the cached tokens per call and the cached price to see the effect. Cache-write fees and lifetimes are not modelled here.
How accurate is the result?
The arithmetic is exact, but real runs vary in length and token counts, and providers add fees this tool does not model. Treat it as a planning estimate and check your provider’s usage report.
Last updated: