Token

A token is the chunk of text a model actually processes. Text is split by a tokenizer into pieces that are usually a word, part of a word, or a punctuation mark, and the model reads and writes these pieces rather than characters. In English prose a token averages around four characters, so a thousand words is roughly 1,300 tokens. Code tokenizes worse: indentation, brackets and symbols each cost their own tokens.

Every limit and every price is in tokens. The context window is a token count. Input tokens and output tokens are billed separately. Your usage limit drains in tokens. That's why a pasted log file or a giant JSON blob can quietly wreck a session: it looks like one message to you, but to the model it's tens of thousands of tokens competing for space.

You don't need to count precisely, but you do need a sense of scale. A typical source file is a few thousand tokens; a test run's output can be more; a whole repo is usually far beyond the window. Load what the task touches, use search instead of reading everything, and trim noisy tool output before it lands in context.