Claude Code Costs, Act II — Where the big hidden costs are

Short summary

Switching between Claude models mid-conversation forces each model to rebuild its own prompt cache, since caches are model-scoped and can't be reused across different models. This creates expensive duplication when bouncing between Sonnet and Haiku, offsetting the cost savings from cheaper models. Measured data shows single-model sessions with persistent warm caching achieve 50–85% cost reductions versus per-turn switching.

• •Prompt caches are model-scoped and can't transfer between Claude models — a architectural consequence of KV cache attention mechanics
• •Model switching causes cache duplication and recurring catch-up token writes on re-entry, offsetting cheaper-model savings
• •Single-model sessions save 50–85% via warm cache reuse; per-turn bouncing incurs 2× write-rate premiums despite warm re-entry

Generated with AI, which can make mistakes.