What Running AI Agents 24/7 Actually Costs

The numbers, plain

Our factory logs every task an agent completes: which agent, which model, how many tokens, priced at published API-equivalent rates (an API is a paid service you call over the internet; its per-token prices are the standard yardstick). From July 10 to September 1, 2026:

One disclosure so the numbers stay honest: most of this ran on a flat subscription plan, so the dollar figures are consumption priced at API-equivalent rates, not an invoice we paid line by line. That is exactly why we frame everything in tokens. The token counts are the ground truth the meter actually recorded; dollars are just one exchange rate applied to them. Manage the tokens and you have managed the bill, whatever plan or provider the tokens ride on.

One more scope note: the tracker covers task-board work. Interactive sessions with the orchestrator (the agent that assigns work to the others) ride the same plan and are not itemized here. And do not read 330 as a success rate: the tracker only writes a row when a task completes, so failed attempts, bounces, and retries never appear in it, and its failure column reads zero no matter how rough the week was. The task board's own history for this period is full of retried and blocked attempts (28 tasks sit blocked as we write this). Those failed runs burned tokens that never reached the meter, which makes every total on this page a floor, not a ceiling.

What does 24/7 actually mean for cost?

The surprise in the data: uptime is nearly free. The factory is available around the clock, but half the calendar days cost nothing at all, and a quiet day of briefings and maintenance runs under $2. Cost tracks shipping, not time. Our three most expensive days ($60, $49, $39) were all days we pushed launches or site overhauls, with dozens of tasks in flight. The meter follows the work, so an idle factory does not quietly drain money the way an idle cloud server does.

It was not always this shaped. Early operations had the bill passing $3 a day and climbing on routine work, which our build log recorded at the time, along with the rebuild that cut the cost three to four times. That rebuild is part of what Level 2 teaches about cost.

Why are the input tokens so low?

Models bill input tokens (what you send: instructions and context) separately from output tokens (what the model writes). Agent loops are notorious for input costs, because a naive loop re-sends its whole history on every step. Our design pushes against that from two sides. First, anything deterministic, meaning same input, same output, every time, no model judgment involved (checking a build, scanning a site, diffing files, verifying a deploy), runs as a plain script that costs zero tokens; agents only get the summary. Second, agents receive lean, prepared context instead of raw logs and full histories. Input tokens are the cost you can engineer away, and we do, everywhere we can.

Output tokens are different. They are the model actually thinking and writing, and you cannot script your way out of them without losing the work itself. That is why we report output tokens as the honest measure of what the factory consumed: 3.37 million over the window, about 10,000 per completed task on average.

What does a single task cost?

These four classes cover 323 of the 330 tasks in the window; the remaining 7 logged without a model label, about $4 between them.

Routing is the lever: the cheapest model that can do the job gets the job. The spread between $0.11 and $1.41 per task is the routing decision, made per task, and it compounds over hundreds of tasks.

Doesn't running local models make it free?

Cheaper, not free, and the accounting is simpler than people make it. You need a computer to run any of this regardless: the orchestrator, the task board, the memory store all live on your hardware whichever models you use. So the hardware and electricity are a cost you were already paying; what hardware you actually need is its own guide. Cloud models add a token bill on top of that baseline; local models do not. The catch is capability, and we wrote up which model sizes actually hold up in agent work separately. In this window, 20 routine tasks ran on a local model at zero marginal cost.

The staff, by name

The agents are named from ReBoot, the 1994 series set inside a computer called Mainframe, which felt right for a factory of programs. The names turned out to fit their jobs.

Off the meter but on the staff: glitch, the orchestrator, named for the do-everything keytool carried by Bob, the show's Guardian hero, and dot and fong, for Dot Matrix who runs Mainframe's operations and Phong who dispenses the advice. Every one of them is a plain file you can read, which is the point of the whole architecture.

Where to learn this

Level 2 teaches the system these numbers come from: the task board, the orchestrator, memory at scale, and the cost monitoring and model routing that keep the meter shaped like this. New to agents entirely? Start with the free primer.