How to cut Claude Code costs

Claude Code bills by tokens: everything the model reads and writes, every turn. Five habits cut most of the waste. The figures below are from our own lab tests at API list prices.

Where the money goes

Each turn sends the whole conversation again: instructions, tool definitions, files read and earlier replies. When that prefix is cached it is cheap to read (on Opus 5.5, $0.20 per million tokens instead of $4); when the cache has expired it is written again at full price. Output is the most expensive part: $20 per million tokens on Opus 5.5.

1. Use a cheaper model for helper work

Searching files and summarising code does not need your biggest model. Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, against $4 and $20 for Opus 5.5. In our lab, moving read-only helper tasks to Haiku cut the total cost per task by 70.9%, with the same correct answers.

2. Ask for shorter replies

Long explanations, recaps and alternatives you did not ask for are output tokens. A short instruction to answer tersely cut the total cost of a Sonnet chat conversation by 25% in our lab, and of an Opus coding task by 4% (code itself should never be shortened).

3. Watch what is loaded into every message

MCP servers put their tool definitions into every request, used or not. A typical setup of 8 plugins and 3 MCP servers costs $0.47 per 100-turn Opus session before any work; a heavy one costs $1.98, or $0.33 when the client loads tool definitions only on demand. Details in what MCP servers cost in context.

4. Don't let a big session's cache go cold if you will come back

Coming back to a 100,000-token session after the cache expired costs $0.4004 to rebuild it; a warm cache costs $0.0204 to read. If you step away often and come back, keeping the cache warm pays. If you rarely come back, it does not: see when Tokensitter won't save you money.

5. Start a fresh session for a new task

A long session carries its whole history into every turn. When the task changes, start a new session (or clear the old one) so you stop paying to re-read work that no longer matters.

Doing this automatically

Tokensitter is a Claude Code plugin that does 1, 2 and 4 for you, shows what 3 costs, and reports what it saved. You can check its savings on your own bill.