What an AI coding agent really costs per developer

There are two ways to pay for a coding agent, and they fail in opposite directions.

Subscription: fixed price, variable ceiling

A monthly plan is predictable to budget and unpredictable to use. You know what it costs; you do not know whether it will last the week. When it does not, the failure lands mid-task and there is nothing to do about it until the window resets.

API keys: variable price, no ceiling

Per-token billing removes the wall and replaces it with an invoice. It is the right model for automation and the wrong one for a person exploring a codebase, because the most expensive turns are the ones where you least want to be thinking about money.

Where the variance actually comes from

Between two developers on the same plan doing similar work, the spread is rarely about volume of work. It is about four habits:

• →Session length. History is re-sent every turn, so a day-long session pays for its own morning all afternoon.
• →Exploration. An agent that has to find the file reads several to find one.
• →Model choice. Running everything on the largest available model is the default, not a decision.
• →Tool output. Unfiltered test and build logs stay in the transcript for the rest of the session.

Working out a team number

For a team, the number that matters is not the subscription line. It is how many people hit their ceiling in a normal week, because that is capacity you paid for and did not get.

$capsul stats

Run it per developer for a week before changing anything. A team that cannot say which of its members runs out on Wednesday cannot tell whether a change helped.

What a tool can and cannot do here

capsul lowers what each turn spends against whatever you are already paying for, subscription or key. It does not raise a limit, it does not add capacity, and it does not replace your plan. It drives the agent CLI you are already signed into.