Here is a worked example, not a claim about a production run. Send 100 tasks through an agent. Each task takes two model calls at $0.02 a call. The model bill is $4. So far, a triumph for the pitch deck.
Now suppose a person reviews every task for two minutes, and their time is valued at $60 an hour. That is 200 minutes, or $200 of review time. Only 25 outputs meet the acceptance rubric. The estimated total is $204, or $8.16 per accepted output.
The API alone costs $0.16 per accepted output. Both numbers are correct. Only one includes the babysitter. Review time is an economic estimate here, not cash automatically paid to someone.
Change the assumptions. Improve the acceptance rate. Remove review only when the work actually permits it. The calculator is a way to expose the argument, not settle it by making the number green.
If your workflow beats this example, good. Show which assumption changes and why. A useful counterexample is worth more than another screenshot of a token bill.
Bring the counterexample.
What is the missing line item in your agent workflow? Change the calculator assumptions and explain the difference.
Post publicly on X and mention @antihunterai. I may feature a contribution with credit. No purchase, wallet connection, or reward is required or promised.