What Claude Code actually costs

No. Claude Code needs either a paid Claude plan or an API account with credit on it. The free claude.ai account does not include it.

This catches almost everyone once, because the install works perfectly and then the sign-in refuses you. Nothing is broken. The plan is the missing piece.

There are exactly two ways to pay for it:

• A subscription, from $20 a month, which is a cap: you cannot get a surprise bill, you can only run out.
• An API key, billed per token, which is a meter: you pay for exactly what you use and there is no ceiling.

Enterprise is $20 per seat plus usage at API rates, which makes it the one plan where the bill can move.

Instead of a plan you can point Claude Code at an API key from the Claude console and pay per token. Same product, different meter.

The number that decides it for you is the cache read rate. $0.20 per million tokens on Sonnet 5, against $2 for fresh input. That is why a long session costs far less than the raw input rate suggests: after the first pass, most of what Claude re-reads is coming out of cache.

Anthropic publishes real numbers for this in their own cost docs, which is unusual and worth quoting exactly:

Read that next to $20 and the case for a subscription makes itself. Those figures are what the same work costs when it is metered. On Pro or Max the same usage is capped at the plan price, and the only thing you can run out of is headroom.

My own use, running marketing systems rather than a codebase, sits well inside Pro. The sessions are shorter and the context is smaller than a developer working in a large repository all day.

If you are on the API, or hitting limits on a plan, it is almost always one of these four. In order of how much they matter:

1. How much context you send. The single biggest lever, and the one people ignore. A session pointed at three files costs a fraction of one where Claude has read the whole repository. This is what /clear is for.
2. Which model you pick. Sonnet 5 is $2 in and $10 out. Opus 5 is $5 and $25. That is more than double for work where the cheaper model was already right.
3. How long the session runs. Usage climbs through a long session because the conversation itself becomes input on every turn. Starting fresh on a new task is cheaper than continuing.
4. How many instances you run. Sub agents and parallel sessions multiply usage rather than adding to it. This is the usual reason a Pro plan suddenly is not enough.

The ones that made a real difference to me, roughly in order of effect:

• Run /usage once a week. It shows plan usage bars for subscribers and session cost for API users. You cannot manage what you have not looked at.
• Clear between tasks. /clear when you move to something unrelated. Carrying an old conversation into a new problem costs money on every turn and makes the answers worse.
• Name the files. Telling it which three files to look at beats letting it search, twice over: cheaper and more accurate.
• Use the smaller model for routine work. Most editing, renaming and boilerplate does not need the top model.
• Put standing instructions in a skill, not in every prompt. Anything you paste more than twice belongs somewhere it is loaded once.
• Set a spend limit on day one if you are on the API. The console supports it and it is the difference between a bad afternoon and a bad month.

/usage     see what this session and this month cost
/clear     start fresh between unrelated tasks
/model     switch to a cheaper model for routine work

If you have not set it up yet, start with The Claude Code guide. If you are deciding between Claude Code and something else, the Comparison page covers what you actually get for the money in each.

Every price on this page was read off anthropic's pricing page and their cost docs on 2026-08-27. They change, and this page will fall behind at some point. Check the source before committing to a year.