Every LLM call I made — split between work and personal use, by model and by harness. Updated weekly.
67% of my LLM usage is work — and 61% of all tokens are served from cache, not recomputed. Most of what I ask has already been asked before.
Lower output:input and rising cache-read share both indicate more reuse and less re-explaining context.
Bars share one scale; the split shows work vs personal within each row.