Output tokens

Generation is sequential: the model produces one token at a time, and each new token requires a full forward pass conditioned on everything before it. That makes output tokens more expensive to serve than input tokens, and provider price lists reflect it with a markedly higher output rate.

Reasoning models add a twist: the internal thinking they perform before answering is billed as output tokens even when the reasoning text is hidden or summarised. A short final answer can therefore carry thousands of billed output tokens behind it, which is why reasoning modes can multiply the cost of a request.

Controlling verbosity (through instructions, max output limits, and structured formats) is a direct cost lever, since every unnecessary sentence in a response is paid at the highest rate on the price sheet.