Total attempts
requested runs x (1 + retry rate)
Counts the extra model and tool calls created by retry loops.
All tools
Free tool
AI Agent Cost Calculator
Go beyond model-call cost. Include platform fees, tool usage, retry loops, acceptance rate, human review, and maintenance to see what a useful outcome actually costs.
This tool runs in your browser. Your workflow data is not uploaded or sent to an AI model.
Working inputs
Replace the examples with real workflow data before using the result.
Your fully loaded cost report will appear here
Include quality and retries so a cheap run does not hide an expensive result.
Open methodology
What this calculator counts, and why.
A low model bill does not mean an agent is inexpensive. Failed outputs, retries, review time, platform fees, and recurring maintenance can become the larger cost. This model makes each assumption visible so you can replace estimates with operating data.
Methodology by Parker JosephUpdated August 12, 2026
Variable cost
total attempts x (model cost + tool cost)
Combines per-attempt model and connected-tool spend.
Human review cost
runs x review share x review time x hourly cost
Converts review minutes into a monthly labor estimate.
Fully loaded monthly cost
variable + platform + review + maintenance
Keeps direct usage and human operating costs in one total.
Cost per accepted outcome
fully loaded monthly cost / accepted outcomes
Divides cost by useful results, not attempted runs.
Use observed inputs
Pull model and tool cost from billing exports. Measure acceptance, review time, retries, and maintenance from run logs over a representative period.
Interpret the result
Cost per accepted outcome is an operating metric, not a promise of ROI. Compare it with the value and risk of the completed task before deciding to scale.
Know the limitations
The estimate excludes taxes, implementation cost, revenue lift, incident impact, and workflow-specific risk. Add those separately when they affect the decision.
Frequently asked questions
Why use accepted outcomes instead of completed runs?
A run can finish and still be unusable. The acceptance rate forces rejected outputs and quality failures into the unit cost instead of hiding them behind activity.
Should every output be reviewed by a person?
That depends on consequence, reversibility, and process maturity. Enter the share of runs you actually review, then use the workflow readiness checker to set an appropriate approval level.
Does this calculator upload my workflow data?
No. The calculation runs in your browser and does not send these inputs to an AI model.