Claude Code usage limits, and how to stop hitting them
The limit is not a bug and it is not really about how much you work. It is about how much context each turn carries.
Claude subscriptions are metered on usage, not on messages. Two people can send the same number of prompts in a week and land in completely different places, because one of them is sending a repository with every question and the other is not.
What the limit is counting
Roughly speaking, the meter runs on tokens processed, weighted by model. That means three things, and the third one surprises people:
- A long session costs more per turn than a short one, because the whole transcript is re-sent each time.
- A larger model costs more per token than a smaller one for identical work.
- Reading files is not free. An agent that opens six files to answer a one-line question has spent a one-line question's worth of your quota six times over.
This is why the limit tends to arrive in the afternoon of a productive day rather than at the end of a slow week. Productivity here means long sessions, and long sessions are quadratic: every turn pays for every turn before it.
Why it always lands mid-task
Because the expensive turns are the ones in the middle. The first turn of a session is cheap. The twentieth carries nineteen turns of transcript, plus every file read along the way, plus every test output. You are never cut off while planning; you are cut off while finishing.
The four habits that actually help
In order of what they are worth over a week:
- Start a fresh session per task. This is the big one. Nothing else on this list comes close.
- Name the file instead of describing it. Exploration is the most expensive thing an agent does.
- Use the smaller model for mechanical work. Renames, formatting, single-file edits.
- Keep loud command output away from the transcript. Filter it before it lands.
Doing it without changing how you work
The habits above are real but they require discipline on every single turn, which is exactly the kind of thing people stop doing by Wednesday. The alternative is to put a ceiling on the context and let it hold automatically.
npm i -g @penra/capsulcapsul ask 'fix the failing auth test' --budget 3000Drives the Claude CLI you are already signed into. No API key, no second bill.
When you have already hit it
There is no way to buy your way out of a session limit mid-window, and there is no flag that resets it. What you can do is make the next window last longer than the last one, which is a question of what each turn carries rather than how many turns you take.
Questions
Why do I hit the limit faster than a colleague on the same plan?
Almost always session length and file exploration. One long-running session that has read thirty files costs far more per turn than ten short sessions that each read two, even when the number of prompts is identical.
Does the weekly limit reset if I change model?
No. The meter is shared. Choosing a smaller model lowers the rate you consume it at, which is a different and more useful thing.
Does capsul need my Anthropic API key?
No. It drives the Claude CLI you are already signed into, so your existing Pro or Max subscription is what answers. API keys are an optional per-session choice, not a requirement.
Will sending less context give me worse answers?
Not automatically, and it is measured rather than assumed. The benchmark reports answer checks next to token counts on every row, so a saving bought at the cost of correctness is visible instead of hidden.
$ npm i -g @penra/capsul