Use accepted work as the denominator
Token cost alone does not tell you what a workflow costs. Include tool and model usage across all attempts, time spent reviewing and repairing outputs, and the fixed costs allocated to this job.
Divide that total by the number of outputs accepted as successful. Keep unsuccessful tasks in the cost numerator. Otherwise a fragile workflow can appear cheap simply because its failures were left out.
Separate observed costs from estimates
Use billing and run records when available. Before launch, use explicit assumptions and label the result an estimate. Record the time window and currency. Do not count a subscription both in per-attempt usage and again as a fixed cost.
The calculator on this site models a monthly set of requested jobs, average attempts per job, and the fraction of jobs ultimately accepted. It is not a prediction of model behavior or a guarantee of savings.
total cost = jobs × attempts per job × cost per attempt
+ jobs × review minutes ÷ 60 × hourly value
+ allocated fixed costs
accepted jobs = jobs × final acceptance rate
cost per accepted job = total cost ÷ accepted jobsCompare equivalent output quality
Compare manual and agent-assisted work for the same accepted deliverable. A draft is not equivalent to a published, checked report. Include the remaining human work in the agent-assisted path.
Time released for other work is not automatically cash saved. Report capacity and cash effects separately. Test lower acceptance and higher review-time assumptions to see whether your conclusion depends on optimistic inputs.
Reduce waste before adding complexity
Inspect repeated retrieval, excessive context, avoidable retries, and outputs that consistently fail review. Narrowing the job or improving the input contract may reduce cost more effectively than introducing another agent.
Measure again after each change. Splitting work into specialist agents can improve isolation, but it can also introduce extra calls and coordination overhead. The right decision comes from your measured workflow.
Sources and further reading
These guides combine linked documentation with our suggested operating practices. Examples are illustrative; they are not customer results or harness benchmarks.
