The total number of tokens consumed by a prompt, including context, affecting inference cost and latency.