Most of what a coding agent spends is context, not answers. These are the kinds of change that cut LLM and coding-agent token use. Your savings depend on your work.
What this means for Claude Code
- Context is the biggest lever for agents. Old tool output stays in the chat and is read again on every later call.
- Thinking costs tokens. Routine work rarely needs deep reasoning. In Claude Code, the effort level controls it.
- Model switches are not free. The prompt cache belongs to one model, so a switch in the middle of a chat writes the whole context again.
- More tools are not free. Every tool definition the model can see rides along in every call.
- Code navigation tools are not a sure win. They can add tokens as well as save them.