Claude Code Cache Miss Costs 12.5x More Than Hit

Claude Code's token bills are shaped less by model pricing and more by cache behavior, according to a developer post that reverse-engineers Anthropic's own prompt caching documentation. The math is straightforward: cache writes cost 1.25x the base rate, cache reads cost 0.1x, making a cache miss 12.5 times more expensive than a hit on the same tokens. The post documents five mid-session actions that silently invalidate the cache and force a full re-read at write prices: switching system prompts, adding MCP servers after a session has started, toggling tool sets, hitting context limits that spawn a new session, and loading new CLAUDE.md files mid-session. None of these trigger warnings in the UI. Essentially: (Anthropic, Claude Code users) are operating with a pricing model where session hygiene determines cost more than usage volume. - Cache writes at 1.25x base, cache reads at 0.1x base -- the 12.5x ratio comes directly from Anthropic's published docs. - MCP server additions and CLAUDE.md loads mid-session are the least obvious triggers, since both feel like configuration rather than session resets. - Developers running long agentic sessions with tool switching are the highest-exposure group. As Claude Code adoption grows among engineering teams, the gap between developers who understand cache invalidation and those who don't will translate directly into measurable cost differences at scale.