Maximizing Value in Claude Code Sessions: Token Efficiency Tips

Anthropic published a practical guide on getting more value from every token spent in Claude Code sessions. It covers the mechanics of token pricing and offers six concrete tactics to reduce waste without switching models.

What Determines Token Cost

You're billed per token, but you're really paying for GPU inference time. Three factors decide the price of a token: model size, input vs output, and caching. Output tokens cost roughly 5x input because decode runs the model once per token. Cached input tokens are cheaper, and prompt caching expires after an hour.

Six Ways to Cut Token Waste

• Run /clear between tasks — prevents irrelevant prior context from being sent to the model, reducing token usage.
• Set your model and effort level before starting — changing them mid-conversation busts the prompt cache, increasing cost.
• @-mention files instead of naming them — the file attaches directly to your message, saving a Read call or a repo search.
• Add quiet flags to noisy commands, or run them in a subagent — command output stays in the conversation for the rest of the session.
• Run /context once in a fresh session — shows what's loaded (CLAUDE.md, MCP tools), so you can cut unnecessary items.
• /compact before you take a break — the prompt cache expires after an hour, and summarizing while cached is cheaper.

The guide emphasizes that being token-efficient isn't about using fewer overall, but making sure the ones you use go toward the actual task. For example, in one session Claude reads the test and the file it covers, edits, and finishes in a few turns. In another, it greps around, reads a dozen files to reach the same two, and drags all that context into every subsequent turn — costing more and making the model think about irrelevant files.