Share of input tokens served from cache rather than processed fresh.
Cached input tokens ÷ total input tokens × 100