How the ai workflow & agent cost calculator works
Estimate the recurring cost of a complete AI task: the model requests, external services, repeat executions and monthly overhead needed to produce an outcome. A task might be one document reviewed or one support enquiry handled; it can involve several separately billed requests.
Method and behaviour
For step j, cost per execution is (input tokens × input rate + output tokens × output rate) ÷ 1,000,000 + the sum of external calls × price per call. Variable monthly cost is initiated tasks × the sum of expected step executions × step cost. Add fixed overhead once.
In the default mode, executions already include retries. In whole-task retry mode, with failure fraction f and maximum total attempts K, expected attempts = 1 + f + … + f^(K−1), final success = 1 − f^K, and step executions per attempt are multiplied by expected attempts. Both successful and failed attempts must have identical expected billable work. Use measured expected executions for branching workflows with different failed-attempt costs.
Expected successful tasks = initiated tasks × final success fraction. Divide the entire total by initiated tasks or expected successes for unit costs. A zero denominator is undefined. Fractional expected completions describe a long-run average, not part of an actual completed task.
Arithmetic uses exact decimal fractions. Display and exported display values round half up to 6 decimal places, with at least 2 for money. Positive amounts below 0.000001 use a threshold label. Changing currency labels does not convert values.
Worked example
Hypothetical CAD example: 100 tasks; a model step with 10,000 input tokens at 1 per million and 2,500 output tokens at 4 per million costs 0.02 per execution. Set it to 3 executions. Add a separate zero-token step with 1 execution and 2 external calls at 0.005 each. At 80% success and no overhead, model charges are 6.00, tool charges 1.00, total 7.00, cost per initiated task 0.07 and per successful task 0.0875. In whole-task retry mode, use per-attempt executions 3 and 1, 20% failure and 3 total attempts: expected attempts are 1.24, success 99.2%, total 8.68 and cost per success 0.0875. Switching modes resets execution fields to 1; re-enter the appropriate counts.
Useful for
- Estimate document-processing workflows.
- Separate model charges from paid searches.
- Compare observed repeat work with independent-attempt assumptions.
- Allocate hosting and monitoring costs across outcomes.
Important limitations
- Use 1–20 steps and at most 20 external charge rows. Monthly tasks are whole numbers up to 1 billion; average tokens up to 1 billion; executions up to 10,000; calls and rates up to 1 million. Decimal quantities allow 6 fractional digits.
- Percentages allow 4 fractional digits, maximum total attempts is an integer 1–20, and overhead allows 2 fractional digits up to 1 trillion. Inputs are at most 64 characters. Use plain non-negative decimals without separators or exponents.
- Taxes, foreign exchange, caching, batch discounts, tiered prices, storage and human review are not inferred. Add applicable charges explicitly or model them separately.
- Independent retries are inappropriate when outages or unchanged invalid input make failures correlated. Neither mode forecasts actual invoices.
Questions specific to this tool
Where do external charges go?
Open External charges inside the step that incurs them. Charges are multiplied by that step's executions. For a once-per-task fee, use a separate zero-token step with one execution.
Why is cost per success different from request price?
It allocates all spending, including unsuccessful attempts and overhead, across expected successes. It is not a provider tariff.
Can I compare CAD and USD?
First convert every rate and overhead amount independently. Selecting a currency only changes labels and clears results.
Publisher: Sitrucanaid Inc. · Implementation updated: 2026-10-17 · Tool version 1.0.0