What one AI agent task actually costs, and why long sessions are the budget killer
All releases
Teams & Billing3 min read

What one AI agent task actually costs, and why long sessions are the budget killer

Plan prices hide the unit that matters: the task. From real production data across hundreds of agents, here is what an agent task costs in machine time, why sessions that run long cost more than they should, and the one ratio that predicts your bill.

PricingAI agentsCosts

Ask what an AI agent costs and you get plan prices: $20 here, $200 there. That answers the wrong question. Agents are not seats, they are workers, and the unit of work is the task. What does one task cost? Almost nobody publishes that number, because most pricing models make it invisible.

We can answer it, because JackHamr bills infrastructure at cost on an itemized bill, and our agents log machine hours separately from tokens. Here is real data from the last 30 days across 363 active agents, with the methodology at the end. Our previous cost post covered the hourly structure; this one covers the task.

What an agent costs per month, for real

Infrastructure cost per active agent (machine only, tokens excluded), last 30 days:

  • Median: $1.81/month
  • p75: $8.19/month
  • p90: $16.06/month

The distribution is the story, not the median. Half of agents barely tick over: a couple of tasks a week on a small machine. But the top decile runs 9x the median, and those are the agents doing real work. If you budget off the average, you will be wrong in both directions.

Note what this excludes: model tokens. An agent's machine cost is the floor. Token spend on top is task-dependent and can dwarf it, which is why bring-your-own-key matters (you see the raw API bill, not a markup).

The two levers that decide a task's cost

A task's cost is roughly machine-hours times machine-size, plus tokens. Two levers move it:

**Session length.** The most expensive task is not the biggest one, it is the one that does not stop. An agent that finishes in 40 minutes on a 4-core machine costs a fraction of one that idles through a 6-hour session, even if both produce the same code. Long sessions usually mean waiting: for a review, for a human answer, for a gate that nobody opened.

**Waiting versus working.** This is the ratio we did not expect to see. Across agents with answered cards, machines spent 1,114 hours waiting on a person against 616 hours actually working. Nearly twice as much time waiting as working. Half of human answers came within 2 minutes (the machine barely paused); the rest stretched out for hours while the meter quietly ran, and about 1 in 5 cards were never answered at all.

That ratio is the single best predictor of your bill. A team that answers agent questions in minutes runs the same workload for a fraction of the cost of a team that batches answers for the afternoon.

What this means in practice

  • **Budget per task, not per seat.** Estimate tasks per week times machine-hours per task. Twenty tasks a week at half an hour each is roughly 10 machine-hours, whatever plan you are on.
  • **Treat unanswered gates as a cost line.** A card that sits for 6 hours is not free. The machine is paused-cheap but tokens and session state still accumulate. Answer gates fast and your cost per task drops without changing anything else.
  • **Small machines finish the same tasks.** Most agent tasks do not need 8 cores. They need to finish. A 2-core machine that completes the task beats a 4-core one that idles.
  • **Long sessions are a smell.** When a task runs for hours, find out whether it is building or waiting. In our data, the wait is usually the answer.

Methodology and honesty notes

Machine cost is measured from 363 agents with real activity in the last 30 days, computed from allocated machine hours at cost. It excludes token spend (we cannot attribute another provider's API bill per task) and slightly overstates cost because paused-time billing predates a fix on Oct 24. The waiting-versus-working ratio comes from 95 agents with answered decision cards over 90 days: 1,114 hours waiting versus 616 hours working. The 1-in-5 never-answered figure applies to cards that waited 7 days before being capped.

If you want the same itemized view of your own agents, that is what our billing is: infrastructure at cost, tokens at provider cost, nothing marked up, nothing bundled.

Try it on your own project.

$30 in free credit, no card needed. Your first agent is running the moment you sign up.

More releases