Input vs Output Tokens

In one sentence

Input tokens are what you send to the model; output tokens are what the model generates—often billed at different rates and driven by different behaviors.

Why it matters

Agent sessions can be input-heavy (big context) or output-heavy (long patches and explanations). Optimizing the wrong side wastes money.

How it works

• Count prompt, files, and tool results as input (subject to product rules).
• Count completions—including code and chain-of-thought style outputs when billed—as output.
• Compare unit prices: output is frequently more expensive per token.
• Reduce input with caching/compaction; reduce output with tighter instructions and less narration when appropriate.

Example

A turn that re-sends a 200k-token repo summary and asks for a three-line fix is input-dominated. A turn that asks “rewrite the module and explain every line” is output-dominated.