Last week, I wrote about prompt cache discipline in Claude Code. I followed the rules by quickly checking the numbers, now I can directly see them.
Release v2.1.151 from Anthropic adds a prompt cache line to /cost (or /usage). It shows the number of requests, hit ratio, the misses, if the cache state is warm or cold and what is the TTL (which was very ambiguous before).
This is more than a cosmetic change.
A cache read is much less expensive than a new input token. A cold session is expensive. When the TTL ends, the tool must write the full context to the cache again. One prompt can then consume a large part of your five-hour limit.
Before this release, you saw only the effect: the limit decreased quickly. The cause stayed hidden. Now you see the cause. If the hit ratio is low, you can clear the session, or you can resume the work before the TTL ends.
Show the metric, and the behavior changes.
Two more numbers I believe could also help in the decision making of starting a new session or continue an old one:
– The current context size in tokens. The hit ratio tells you the cost of each turn. The context size tells you when to compact or to hand off. My limit is 150k tokens. Easy go get with /context all, not directly seen.
– The growth of the token usage between the last checkpoint and now. A total is a history. A delta is a signal. It shows you which step made the cost increase. A value for each model would show where the expensive work occurs.
Which other indicators would you like to see in your status line?
#ClaudeCode #Anthropic #LLM
