A four-state LLM spend card: stacked bars split every period's tokens into input, cached input, output and reasoning, a cost line derived in-component from per-million rates rides a second axis sharing one gridline set, and the cache-hit share plus the money the cache spread saved are called out above it.