Models do not read words. They read tokens, which are chunks of text, often a few characters or part of a word. In English a token is roughly three quarters of a word on average, and other languages and code often need more tokens for the same content. Images and audio are also converted into tokens.
Input tokens include everything sent in the request: the instructions, the conversation so far, any documents or search results you attach, tool outputs and the new question. In a chat, the earlier messages are usually sent again with each turn, so the input grows as the conversation continues.
Input is limited by the context window and is billed per token. Some providers offer lower prices for cached input, which is a prompt prefix that was used recently.
Keep prompts focused, send only the documents needed, trim old history and use caching for large instructions that repeat.