Optimizing Claude Code Token Costs: Cache Hit Rates and Session Management Strategies

Claude Code resends the entire system prompt, project context, and conversation history with every request, so costs are proportional to the session size. The actual unit price is determined by the prompt caching hit rate, with cache reads charged at approximately 10% of the standard input rate.

Cost Structure and Measurement Tools

Requests consist of three layers: system prompt, project context, and conversation history. Cache writes cost 2.0x (1-hour TTL) or 1.25x (5-minute TTL) the standard rate, but this is a one-time investment that allows for reuse at lower rates in subsequent turns. Use the /usage command to check token statistics and estimated costs, and /context to inspect the context footprint of MCP tool definitions, CLAUDE.md, and other elements. /insights analyzes repetitive tasks and friction points, providing an HTML report.

Session Management Strategies

To improve cost efficiency, use /clear to reset history when switching to unrelated tasks, and use /compact to summarize context when continuing the same task. Compaction should be executed while the cache is maintained to keep costs low; if the direction is wrong, /rewind, which leverages the existing cache as is, is more cost-effective than compaction.