The model does not see words, it sees tokens: for a common word one token is the whole word, for rarer ones the word falls into several pieces. In English roughly four characters make a token. Everything you pay for and every limit you run into is measured in tokens — the input, the output and the context window alike.
This has a direct consequence for every language other than English: tokenisers were built mainly on English text, because that is what there was most of in the training material. The same document in another language burns more tokens, so it costs more and fills the window sooner than the English version does.
The most common misconception is that token is a synonym for word. It is not, and the gap is widest exactly where it matters: proper nouns, technical terms, numbers and non-English languages. Anyone estimating in words underestimates the cost outside English, every time.
How tokens and the window hang together is set out in Tokens and the Context Window: Why Length Costs Money, a post of its own. Directly related terms: tokenizer, context window and inference.