A model cost chart without cached-token and output-token columns is hard to trust

A note on why AI cost reports should split input, cached input, output, latency, and task type.

When an AI feature gets expensive, a single total-token column does not explain what to fix. If uncached input is large, the prompt shape may need work. If cached input is large, the cost may already be partly optimized. If output tokens dominate, the response format is probably too loose. If latency is high, a tool call or external service may be the real issue. I now prefer cost rows with provider, model, prompt version, input, cached input, output, latency, and task type. Otherwise a cost chart becomes a guess with numbers.