Switching models no longer reuses a stale prompt-cache record

The prompt cache lets the service reuse work from earlier in a conversation, so repeated parts of a request do not have to be processed again. Claude Code keeps a record that links one request to the next for this purpose.

Claude Code now counts model switches. When you change to a different model, that pending record is set aside or dropped. Before, the record was only released when the chain ended, when the account changed, or when caching was not allowed. A model switch is now a fourth reason. The check also takes into account whether the session was resumed, meaning picked up again after being closed.