Three ways to run AI coding agents

Agent Factory Floor

Three factory lines side by side. Copper robots work on our Mac, blue robots on Claude in the cloud, and green robots on GPT in the cloud. Boxes ride each belt to a gate that sends them to Done or back for rework.

Robot
One AI coding agent, a program that writes code on its own. Each robot on the floor is one agent.
Line
One way of running robots: our own Mac, Claude in the cloud, or GPT in the cloud.
Job
One piece of coding work, shown as a box on the belt.
Good job
A box that passes the check at the gate. The rest go back for rework.
Robot-hour
One robot working for one hour.
Account
One paid subscription. It gives a set amount of robot-hours each week, like a battery that refills every week.
Tokens
AI reads and writes in tokens, pieces of words. 1,000 tokens is about 750 words.

Pick how many robots you want. Then watch what each line does with the same order.

Order desk

8robots
More settings

Factory floor

All three lines get the same order. Each box on a belt is one job. The gate checks it: green goes to the Done bin, red slides back for rework.

Line 1

Our Mac

Every robot here plugs into one Mac Studio. Each new robot shares the same machine, so every robot slows down, and when the memory shelf is full no more can start.

In this picture: Done 0 Rework 0Spent $0Each good box —

Line 2

Claude Opus 5.5

Model

In this picture: Done 0 Rework 0Spent $0Each good box —

Line 3

GPT

In this picture: Done 0 Rework 0Spent $0Each good box —

What a month costs, and why

What one good job costs

Every box costs money to make, even the ones that fail the check. So each good box also pays for the failed ones.

Where each line breaks

Each chart runs from 1 to 64 robots. The dashed line marks your order. Red marks show where a line stops growing.

    Our Mac up close: jobs done a day

    Good jobs a day, all three lines

    Cost a month

    Cost per good job

    Report card: real coding tests

    Each test gives agents real coding work and counts how often they get it right. Bars run from 0 to 100%. The badge says who ran the test: the model's maker, or outside testers.

    How our subscriptions work

    We don't pay for each token. We pay a flat price for accounts, and each account gives a set amount of robot-hours every week. Here is how that works, step by step.

    One robot-hour, priced every way

    What one robot working for one hour costs. Our subscription prices assume every account gets used up.

      How big companies buy

      Change the numbers

      Each number below comes from a source or from our own tests. Change any of them and the whole page follows: the floor, the bills and the charts.

      Good at, bad at

      Each way of running agents has real strengths. Numbers here follow the settings above.

      How we measured, and sources

      Each number on this page has a tag like [M1] or [S4]. Tap a tag to jump to its entry here. M means we measured it ourselves.

      What we measured

        Sources