Claude Code · Token cost playbook

Spend fewer tokens, get more done.

You're billed for every token Claude processes — and your whole conversation is resent on every turn. Shrink the context, right-size the model, and stop re-paying for stale history. The biggest levers are up top; everything else is tagged by impact.

The mental model
context tokensprompt + files + history + tools × every turnhistory resent each time ≈ your billgrows as the thread grows
A 50-turn session doesn't bill 50 prompts — it bills 50 growing copies of the conversation. That's why context hygiene beats everything else.
≈90%
off cached input — automatic, nothing to turn on
10×
price spread Haiku → Fable — routing is the single biggest lever
≈7×
tokens an agent team burns vs. one session
Nothing matched.Try a shorter or different term - like “compact”, “model”, “mcp” or “subagent”.