How do hosted model APIs charge you?

Hosted model APIs do not charge per request or per question. They charge per token — the small chunks of text a model reads and writes. A token is roughly three-quarters of a word in English, so a paragraph might be 100 tokens and a long document might be tens of thousands.

The bill splits into two parts that are priced separately: the tokens you send in (input) and the tokens the model writes back (output). Rates are quoted per million tokens, and often by context length too: on OpenAI’s pricing page, for example, GPT-5.5 lists $5.00 for input and $30.00 for output per million tokens for requests under 272K tokens, rising to $10.00 input and $45.00 output past that threshold.^[1] The two numbers are almost always different, and output is the larger one — and on a model with tiered pricing like this, which tier you land in changes the bill as much as the token count does.

The reason output costs more comes from how the model produces it. Input is read in a single pass, so the whole prompt can be processed at once. Output has to be generated one token at a time, because each new token depends on every token before it. Writing a 1,000-token answer means roughly 1,000 sequential steps, while reading a 1,000-token prompt is one. That gap in work is what the price difference reflects.

The practical takeaway: a long answer can cost more than a long question, even when the question has far more words. If you want to control spend, the levers are the size of what you send, the length of what you ask for, and the features that cut the cost of repeated context.

The ratio between input and output pricing is not fixed across providers, models, or even tiers of the same model — GPT-5.5’s roughly 6x gap in its short-context tier narrows to 4.5x once a request crosses into the long-context tier, and that’s one data point, not a law. What stays consistent is the direction: output pricing tracks the sequential-generation cost, which is a property of autoregressive decoding itself, not a choice any one company made, so you should expect it to price higher than input on any provider using the same kind of model.