Stacked by model. Dollars at public API prices, not a subscription invoice.
The same days by volume. Where this disagrees with the cost chart is the finding.
Input, output, cache write and cache read per model — cache reads explain the headline volume.
Every machine pushes its own transcripts. This split ignores the filter above.
Pushed hourly from every machine running Claude Code — ccusage reads its local
transcripts, so each machine reports only its own and work done on claude.ai/code is nowhere.
A machine that never pushes is simply missing from the totals. Costs are
ccusage's own calculation against public API rates: on a Max or Pro plan the
figure is what this would have cost via the API. Yesterday and today are re-pushed every
hour, so today's row keeps growing until the day ends.