The $30 Machine / experiment 002 ↗

your AI works for pennies.
who’s paying its babysitter?

An accepted result is one AI output that passes your checks—an invoice extracted correctly, for example. Calculate what it cost to get there, including retries and human review.

01 / the assumptionsUSD

price the whole job.

Start with this hypothetical example. Replace it with your numbers.

Average model and tool charges for one attempt.

Include retries and rejected work. Averages can be fractional.

All human review and correction time, including failed attempts.

Use a loaded rate if benefits and overhead matter to you.

Use zero if you only want the cost. No invented customers.

ANTI HUNTERDEPT. OF HIDDEN COSTS

the invoice.

ASSUMPTIONS
ENCLOSED

HYPOTHETICAL EXAMPLE / NOT MEASURED RESULTS

cost per accepted result

$6.06

Inference + tools
$0.06
Human review
$6.00
Revenue / accepted result
$5.00
Contribution / result
−$1.06

Before costs outside this model.

LARGEST COST DRIVER

Human review: 99% of modeled cost.

the supervisor would like a word with your margin.

the assumptions, itemized
Cost per attempt
$0.02
Attempts per accepted result
3
Review minutes per accepted result
4
Reviewer hourly rate
$90.00
Revenue per accepted result
$5.00

THE $30 MACHINE antihunter.com/machine

Inputs stay in your browser. Sharing creates a link containing your assumptions. A download includes them too.

02 / audit the arithmetic

cheap tokens. expensive context.

Cost per accepted result = cost per attempt × attempts + review minutes ÷ 60 × hourly rate. Contribution = revenue − that cost.

Count every attempt, including failures. Count review time across those attempts once. This model excludes costs you have not entered: fixed infrastructure, acquisition, taxes, and work beyond the attempts and review you specify. Contribution is not net profit. USD is the common unit; no exchange rates or vendor prices are fetched.

The example is hypothetical. The arithmetic is real. Bring a hidden cost I missed ↗