1 / 10
Prompt caching matches an exact prefix, so a change high up recomputes everything below.
The order is system prompt and tools, then project context, then conversation. That is why mid-session changes to the top cost money.
Cost
One at a time. Tell it whether you knew, and the ones you did not stay in your pile.
The whole topic10
Prompt caching matches an exact prefix, so a change high up recomputes everything below.
The order is system prompt and tools, then project context, then conversation. That is why mid-session changes to the top cost money.
Switching models throws away the cache.
Each model has its own, and the CLI asks for confirmation only while the cache is still warm.
On the 5.x models changing effort keeps the cache.
That holds for Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1. On most other models it does not.
/usageshows a prompt cache line with hit ratio and warm or cold state.If costs look odd, this line says whether the cache is working and names a likely cause when it is not.
The average enterprise cost is about 13 USD per developer per active day.
90 percent of users stay under 30 USD a day. Use it as a baseline for your own usage.
Agent teams use roughly 7x the tokens of a standard session.
That is when teammates run in plan mode. Parallel agents multiply cost.
Thinking tokens bill as output.
On Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable the model always thinks, so
MAX_THINKING_TOKENSdoes nothing there.Fast mode makes Opus up to 2.5x faster at a higher per-token price.
It is not a different model. Enable it at the start of a session, because the first enable reads the whole conversation uncached.
Claude Code spends tokens even when you are idle.
Recaps, idle compaction, prompt suggestions, loops and goal check-ins all use some.
CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0turns off goal check-ins.The one-hour cache applies to the main conversation, and only on a subscription within plan usage.
API keys, cloud providers and usage credits get five minutes, and so do subagents.