What an always-on agent really costs

There are two bills. One is fixed and boring. The other is variable and is the reason people write posts titled "how I spent $400 on an AI agent in a weekend." Almost everything useful to know about agent cost is about keeping those two apart.

Bill one: hosting

A fixed monthly price for a machine. Here are the actual plans:

That is the whole of bill one. It does not move with usage, the first agent month is free on the micro-VM plans, and the machine stops when idle so you are not paying for something to sit there overnight. The dedicated multi-agent boxes are excluded from the free first month, because a month of a dedicated 8–16 GB machine is a real hardware cost and we would rather say that up front than quietly not honour it.

Bill two: model tokens

Every message, every tool call, every scheduled run sends tokens to a model provider. Those are forwarded at provider cost with no markup — hosting is the paid product here, not a margin on your tokens — or you bring your own API key and pay your provider directly.

This is the bill that varies, and it varies by a lot more than people expect. The same $9 box can cost you two dollars a month in tokens or two hundred, and the difference is almost never "how many messages did I send."

What actually drives the token bill

Whether anything runs on a schedule

The single biggest factor. A chat agent costs what you use it for. A scheduled agent costs the same amount every day whether or not that day had anything in it. An hourly job is twenty-four times a daily one for roughly the same information — see scheduled agents.

Whether the agent uses tools

A conversational answer is one model call. A coding agent reading files, running a command, reading the output and trying again is a dozen — each carrying the accumulated context of the ones before it. This is why the hosted coding agents have a different cost profile from a chat bot on identical hardware.

Whether it retries

Here is the failure mode that produces the horror stories. An agent that cannot complete something and keeps trying does not crash and does not alert. It works, expensively, in a loop, at three in the morning. A stuck coding agent is the most expensive single thing that happens on this platform, and it is completely silent unless something is watching for it.

Which model you picked

Frontier models cost multiples of the small ones. For triage, classification and routing, the small one is genuinely fine, and paying frontier prices for "is this urgent, yes or no" is the most common avoidable waste. This is a real argument for Hermes on multi-step work: different steps can use different models.

A worked example

A Telegram assistant that answers maybe thirty messages a day and sends one morning digest, on a mid-range model:

• Hosting: $9–$19 a month, fixed, depending on whether it needs the fuller runtime.
• Conversation: low single-digit dollars a month. Thirty short exchanges a day is not much work for a model.
• The daily digest: usually more than the conversation, because it reads everything from the day before summarising it.

So: roughly the cost of a couple of coffees, and the fixed half dominates. Now change one thing — make the digest hourly and have it read the same sources every time — and the variable half is suddenly the larger number by an order of magnitude. Nothing about that change looks expensive when you make it. That is the whole problem.

How to not get surprised

1. Look at the first week and multiply by four. If that number is uncomfortable, fix it in week one, not after the invoice.
2. Watch cost per conversation, not the total. The total tells you that you spent money. Cost per conversation tells you which conversation to go and read. That drill-down is on the Billing tab.
3. Start daily, not hourly. Increase frequency only when you can point at a run that should have happened sooner.
4. Give every agent a give-up instruction. "If you cannot do this in a few attempts, stop and tell me" is one sentence and it is the difference between a failed task and a runaway one.
5. Keep a stop button within reach. Agents here report to ClawMetry, which shows live spend per agent and can pause or kill one mid-run. For an unattended process, the ability to stop it is a cost control.

When credits run out

The agent is stopped, not silently billed. Top up and it resumes where it left off; volumes and history persist. This is deliberate — a runaway agent should hit a wall, not an overdraft.

The comparison worth making

The number that matters is not what the agent costs. It is what the agent costs against the thing it replaces. An agent that handles the routine majority of a support inbox for the price of a subscription is straightforwardly worth it. An agent that generates a report nobody reads is not worth anything at all, and that is true at $2 a month as much as at $200.