Claude Code Test Exposes Cache-Switching Costs
A 27-run test of Claude Code 2.1.302 shows that the agent resends its growing context each turn, with prompt caching doing the cost-control work. Switching models or returning after cache expiry forces a cold re-read, making cheaper model routing unexpectedly expensive.
Claude Code’s cache turns model choice into a workflow decision: switching models mid-session can erase the savings you expected. The test highlights a cost issue developers rarely see in the interface.
- –Every turn carries forward system instructions, project context, conversation history, and tool results.
- –Model switches use separate caches, forcing the next request to reread the full conversation.
- –Cache expiry after inactivity creates another expensive cold-start turn.
- –Developers should choose a model early, use `/clear` between unrelated tasks, and monitor cache reads versus cache creation.
- –Results will vary with context length, authentication method, TTL settings, and model pricing, so the test is a warning rather than a universal cost estimate.
DISCOVERED
1h ago
2026-10-18
PUBLISHED
1h ago
2026-10-18
RELEVANCE
AUTHOR
DIY Smart Code