Token costs

1 · What it is

Apps often reach an AI model through an API, a service they call over the internet. When you do this, you pay for tokens. A token is a small piece of text, such as a short word or part of a long word. In Gemini, one token works out to roughly four characters. The bill counts tokens both ways: the text you send in (input) and the text the model writes back (output).

Those two directions have different prices. In one Anthropic price table, output costs five times as much as input. Reasoning models, which work through a problem step by step, add a twist. They think in hidden tokens before answering, and those are billed as output. Gemini’s price page says the same thing: its output price includes thinking tokens.

Picture a homework helper app that sends the same long set of rules with every question. That repeated start can be cached, which means stored for reuse. Saving it to the cache costs a little more than normal input: 1.25 times the base price for a five-minute cache. Reading it back costs about a tenth of the normal input price on many Claude models. So caching pays off after one reuse. Some jobs can wait a day for an answer. Those can go through a batch service, which runs them later, at half price. Tools such as web search can add their own charges on top.