Why “Word Count” Is the Wrong Mental Model
When you type a message to an AI model, it doesn’t see words the way you do. It sees tokens, small chunks of text produced by a process called tokenization. A token might be a whole common word like “the,” part of a longer word like “token” splitting into “tok” and “en,” or even a single punctuation mark. As a rough rule of thumb, one token is about four characters of English text, or roughly three-quarters of a word. A 500-word blog post is usually somewhere around 650 to 700 tokens, not 500.
Why This Matters for Pricing
Every AI API charges based on tokens processed, both the tokens you send in (the input or “prompt”) and the tokens the model generates back (the output or “completion”). This is why pasting a long document into an AI chat and asking a short question can cost noticeably more than asking the same question with no document attached: you’re paying for every token of that document to be read, even though your actual question was one sentence. Output tokens are typically priced higher than input tokens, which is why a model that writes a long, verbose answer costs more than one that answers just as accurately in three sentences.
Why Odd Words Confuse Models More
Common English words usually map to a single token because they appear so often in the training data that the tokenizer learned them as one unit. Rare words, made-up names, non-English text, or unusual number formats often get chopped into several smaller, less intuitive tokens. This is part of why AI models are sometimes worse at tasks like counting the letters in an unusual word, or doing arithmetic with long numbers: the model isn’t seeing “137,482” as a clean number, it may be seeing it as two or three separate token fragments, making digit-level reasoning genuinely harder for it than for you.
Why Context Windows Are Measured in Tokens
When you hear that a model has a “200,000 token context window,” that’s the maximum combined size of everything the model can consider at once: your entire conversation history, any documents you’ve shared, its own previous replies, and the new output it’s about to generate. Once a conversation’s token count approaches that ceiling, older content typically gets dropped or summarized, which is a major reason long AI conversations seem to “forget” earlier details even though nothing was technically deleted from the chat log you can see.
A Practical Takeaway
If you’re trying to control AI costs or keep a long project within a model’s memory, think in tokens, not words or pages. Trimming unnecessary repetition from a prompt, summarizing a long document before pasting it in, and asking for concise rather than exhaustive answers all directly reduce token usage. And if a model seems to be struggling with an unusual word, a long number, or a very specific technical term, it may genuinely be a tokenization issue rather than a reasoning failure, worth rephrasing or spelling it out differently rather than assuming the model is simply wrong.
