How Much Does It Cost to Run an AI Agent?

Why there are two bills

A self-hosted agent separates two things that a SaaS agent bundles into one subscription. That separation is the whole point of bring-your-own-key, and it is also why the question "what does an agent cost" has no single answer.

• The server. A fixed monthly amount for the instance the agent runs on. Predictable, small, and unaffected by how hard you work the agent.
• The model. Billed by your AI provider, on your own account, in proportion to how much the agent thinks. We never touch it, cannot mark it up, and do not see it.

The practical consequence is that you can size the first exactly and only estimate the second. Most of this page is about estimating the second honestly, because that is where people get surprised.

What the server costs

Both include the first week free, are billed month to month with no contract, and are dedicated resources rather than a share of something busier. Because the agent orchestrates rather than infers, it spends most of its life idle waiting on the network, which is why these numbers are so much lower than people expect. There is no GPU in the list because a bring-your-own-key agent has no use for one.

The honest comparison here is not against zero. A VPS you administer yourself has a similar sticker price and then adds the parts nobody quotes: the operating system, the reverse proxy, certificate renewal, patching, backups and being the person who notices at 3am. Those are real costs paid in a currency that does not appear on a card statement.

What drives the model bill

Providers price by tokens consumed, and an agent consumes them differently from a chatbot. This is the single most useful thing to understand before you estimate anything, because it explains why agent costs surprise people who are used to chat pricing.

A chat assistant makes one model call per message you send. An agent makes one call per step of its reasoning loop: read the task, decide, call a tool, read the result, decide again. A task that takes eight steps is eight calls, not one. Multiply that across a day of scheduled work and the arithmetic moves quickly.

Worse, each of those calls usually carries the whole conversation so far. Step eight re-sends everything from steps one through seven so the model knows where it is. That makes the cost of a long task grow faster than the number of steps, which is the mechanism behind most unexpectedly large bills.

• Steps per task. A one-shot summary is cheap. A research task that browses, reads, cross-checks and writes is not.
• Context carried per step. Feeding an agent a large document means paying for that document again on every subsequent step of the same task.
• How often it runs. A task that fires every ten minutes runs roughly 4,300 times a month. The per-run cost matters far less than the frequency you set.
• Which model. Capability and price vary by more than an order of magnitude between models, and the difference matters most for agents because a weaker model retries more.

A way to estimate before you commit

You do not have to guess. The reliable method is to run one representative task, look at what it actually consumed on your provider dashboard, and multiply by how often you intend to run it. That single measurement is worth more than any published estimate, because your tasks are not anyone else’s.

1. Define one real task, the kind you actually want automated.
2. Run it once and note the usage your provider records for it.
3. Multiply by the frequency you plan. A task run hourly is roughly 720 times a month; every ten minutes is roughly 4,300.
4. Add the server cost. Compare the total against what the task is worth to you in time saved.

Do this before automating anything on a short schedule. Frequency is the variable people set casually and it is the one that dominates the bill.

Can you run an agent on a free tier?

Partly, and it is worth knowing exactly where the wall is rather than discovering it mid-task. OpenRouter, a common choice for Hermes because it fronts many providers behind one key, publishes limits on its free model variants: 20 requests per minute, and 50 requests per day until the account has purchased $10 of credits at some point, after which the daily allowance rises to 1000.

Read that through the lens of the previous section. Fifty requests a day is not fifty tasks, it is fifty reasoning steps, which a single research run can consume on its own. This is why agents on free tiers characteristically complete short tasks and die halfway through long ones, with no obvious error.

A free tier is a genuinely good way to evaluate an agent and a poor way to run one in production. If you are testing, it costs nothing. If you are automating something you depend on, the daily cap is a reliability problem before it is a cost problem.

Keeping the model bill down

Four levers, roughly in order of how much they move the number:

• Run less often. Reducing a schedule from every ten minutes to hourly cuts the bill by about six times and is frequently invisible in outcome. This is the biggest lever and the most commonly overlooked.
• Keep tasks short. Long-running tasks pay for their own history repeatedly. Several small tasks usually cost less than one large one that does the same work.
• Feed it less. Pass the relevant extract rather than the whole document. Context is charged on every step that carries it, so trimming once saves many times.
• Match the model to the job. The cheapest model is a false economy when it retries, and the most capable one is waste on a task that only needs to reformat text. Different tasks can use different models.

Notice that none of these is about the server. The instance is a fixed few dollars and stays there; everything worth optimising is on the model side.

How this compares to a SaaS agent

A hosted agent product charges one subscription and absorbs the model cost into it, which is genuinely simpler. What you give up is visibility and control: you cannot see what a task cost, you cannot choose a cheaper model for a simple job, and you pay the same whether you run one task a week or a hundred a day.

Self-hosting inverts that. Two bills instead of one, more to understand, and in exchange the usage cost is proportional to what you actually do, the keys and data stay yours, and nothing prevents you from moving. For light use the total is usually lower. For heavy use it depends entirely on how well you apply the four levers above.