Claude Code's Prompt Caching Mechanism and Cost Optimization

Claude Code uses Prefix Caching to cache previous conversation content. When the beginning of the conversation (system prompt, project context, conversation history, etc.) matches, that portion is processed at the low price of about 10% (0.1x) of the existing input cost.

The key factors that determine caching efficiency are as follows:

• Prefix match: If the previous context is modified in the middle of the conversation, a Cache Miss occurs, and in this case the full context is billed again.
• TTL (Time To Live): The cache retention time is set to 5 minutes or 1 hour depending on the authentication method. When the 5-minute TTL is set, if the conversation flow is interrupted for more than 5 minutes, the full context cost is incurred again on the next turn, so caution is needed.
• Cache cost structure:
  • Cache Read: about 10% (when using an existing cache)
  • Cache Write (5m TTL): 1.25x
  • Cache Write (1h TTL): 2x

For active sessions, the TTL is refreshed each time a Cache Hit occurs, maintaining the cache state.