Spending limits and effort checkpoints
Build-along · ~8 min at your own keyboard
Long autonomous runs can spend real money; one course author reported an agent-team audit costing about $80 in an afternoon. There are two kinds of limit. Hard limits are enforced outside the model: your subscription's usage limits, a spend limit set in the API console if you pay per use, or a turn limit on a scripted run. A line in your prompt ('stop after 20 pages') is a checkpoint request — useful, but the model can miss it. Use both, and know which is which.
The long version: Watching cost and usage ↗
Do this
0/4Checkpoints
- ▸You should now see a dollar or usage figure after /cost.
- ▸You should now see where your hard limit is set, written down.
- ▸You should now see whether Claude stopped at your checkpoint, and at what count.
The three ways this goes wrong
- /cost shows nothing useful on a subscription plan. — On Pro or Max, usage limits replace dollars. Use the usage view to see how much of your limit you spent.
- The task ignores your checkpoint line. — That's why it's not a cap. Make it countable ('stop after 20 pages' beats 'be careful'), put it at the start and end of the prompt, and rely on your hard limit as the backstop.
- A long run burns through your limit before you notice. — Start long runs with a smaller test. Check /cost halfway.
Dated facts in this lesson: Hard limits for Claude Code are the subscription plan's usage limits, workspace spend limits in the Claude Console for API use, and --max-turns on headless runs (researched Oct 2, 2026, not yet confirmed by a person: check the source before you rely on it).
Share what happened
Paste your usage reading, where your hard limit is set and its value, your in-prompt checkpoint line, and whether it fired (and at what count).
Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].
🔒 Don't paste confidential, personal-health, or payment info.