Cost
Output tokens
Output tokens are what the model generates. They typically cost four to six times more than input tokens.
Also written: completion tokens
Generation is sequential: each token requires a full forward pass over the model, where the entire input is processed in one. The pricing reflects that compute difference directly.
It also means output length drives latency almost linearly. Asking for a shorter answer is simultaneously the cheapest and the fastest optimisation available.
In practice
Output typically costs three to five times input on the same model, which inverts the intuition that long prompts are the expensive part. A 20,000-token document summarised into 800 tokens costs more for the 800 than a lot of people expect. "Answer in at most three sentences" is a cost control instruction, not a style preference.
Common questions
How do I limit output tokens?
Two mechanisms, and you want both. Set max_tokens from the longest legitimate answer as a hard ceiling, and ask for the length you want in the prompt so the model aims for it rather than being cut off at it.
Why are output tokens more expensive?
They are generated one at a time, each requiring a full forward pass, while input is processed in parallel during prefill. The price difference reflects a real difference in compute per token.