What Claude Code really costs to run

Claude Code is included with the consumer Claude subscriptions, and it also runs against an API key billed per token. That is the whole choice, and it is worth making deliberately rather than by whichever you happened to set up first.

A subscription is a flat monthly fee with usage limits attached. An API key is metered: you pay for what the agent reads and writes, with no ceiling except the one you configure. The tiers and the rates both change often enough that the only sensible move is to check Anthropic’s pricing page the day you decide.

One thing that catches people out: the free Claude.ai tier does not include Claude Code. If you are evaluating it, you are choosing between a paid subscription and a metered key, not between free and paid.

Subscription usage is measured in rolling windows — a short one that resets through the day, and a longer weekly one. Hit either and the agent pauses until the window resets, which is a different and much less alarming failure than a bill that keeps climbing.

The practical consequence is a rhythm, not a number: heavy days feel fine until they suddenly do not, usually in the middle of the one task you cared about. People who work this way learn to spend the expensive part of the window on the expensive part of the job.

Almost all of it is context. Every turn re-sends the conversation and whatever the agent has read, so cost grows with how much the agent is holding, not with how many messages you send.

Which means the levers are the boring ones. Point it at the three files that matter instead of letting it explore the repository. Start a fresh session when the subject changes rather than carrying an hour of unrelated history into it. Keep the acceptance criteria in the request — a task that is redone twice costs three times, and vagueness is what causes the redo.

The box. If you have moved the agent off your laptop so it keeps working, something is now running around the clock. A VPS sized properly for this — 4 GB of RAM is the honest floor, more if it drives a browser — is a real monthly line, and the cheap tiers that look tempting are the ones that fall over under a build.

Your hours. The install is an evening. Keeping it alive is the recurring one: the reboot that did not bring it back, the session that died with the SSH connection, the credential that expired in the middle of a job. Price that at whatever your hour is worth and it usually dwarfs the hosting line.

The second subscription. Self-hosted agents are bring-your-own-model. A $4 host is $4 plus whatever the agent spends thinking, and for anything genuinely useful the second number is much the larger. Compare total monthly cost, or you will pick the cheap box and be surprised by the invoice that matters.

Metered API billing wins if you use the agent in short, sharp bursts — a few focused tasks a week, each one scoped. It is genuinely cheap at that shape, and the cap you set is the only protection you need.

A subscription wins the moment you use it daily, because the thing you are really buying is the absence of a meter running while you think. People who watch a counter use the tool less, and using it less is the expensive outcome.

And if the agent needs to be awake when you are not, add the box and the hours to whichever column you were in. That is the comparison that usually surprises people — the VPS runbook is what it takes in practice, written out in full.

Our plans run from $20 to $200 a month, and both of the bills above are inside them: your agents run on a dedicated Claude subscription we buy and hold for you, and each agent’s machine is included. There is nothing of yours to connect and no API key to buy.

The trade is control. You do not choose the model version, you do not tune the setup, and your capacity is what the subscription behind your plan holds rather than an unlimited meter. If picking your own model matters to you, an API key and a VPS is the honest answer and it is a good one. The full ladder is on the pricing page, with what each rung includes.