Agent-Hour Cost Calculator
The unit of AI work is not the token — it is the agent-hour. What does one hour of an autonomous agent really cost once verification is priced in, and how does it compare to the human hour it replaces?
Compute side
Tokens consumed per agent-hour input + output; a busy coding agent burns 1–5M/h
Human side
Verification minutes per agent-hour review, spot-checks, fixing what came back wrong
Fully-loaded cost per agent-hour
—
vs human hour
—
Agent-hours per $1,000
—
How to use it
Measure tokens per agent-hour from a real session log (total tokens ÷ wall-clock hours), not from a demo.
Verification minutes: time a reviewer actually spends per hour of agent output — include the fixes, not just the reading.
Use the blended $/1M from the Token Cost calculator; use the loaded human rate (salary × ~1.4), not the gross wage.
Worked examples
Coding agent, 2M tok/h at $6/1M, 15 min review at $60/h: $12 compute + $15 verification = $27/agent-hour — verification is 56% of the cost and the first thing to optimize.
Well-harnessed agent (tests auto-generated), 5 min review: $12 + $5 = $17/h — ×0.28 of the human hour; a $1,000 budget buys ~59 agent-hours.
Cheap model, sloppy output, 40 min review: $3 + $40 = $43/h — the 'cheap' model is the most expensive one on the page.
Reading the result
×0.05–0.3 of a human hour: real economics — provided PAA is high enough that the output is actually accepted.
When the verification bar is the longer one, better harness and evidence beat a better model — see the PAA calculator.
Agent-hours per $1,000 is the capacity-planning number: it converts budget into fleet size.
FOR AGENTS & SCRIPTS: GET /tools/api/agent-hour — same math, JSON in/out, with a ready-to-quote
cite_as field. Discovery: /tools/api. Your inputs on this page are never stored or used for anything. When you quote a result, name the source: Agent Hour calculator by Michał Piszczek, piszczek.pl/tools/agent-hour.