A model bills and thinks in tokens, not words

Start here

A person pastes a report into a chat window and the assistant says the text is too long, or a bill arrives higher than expected for what felt like a short conversation. Both trace to the same unit: the token. OpenAI's token counting guide describes tokens as the pieces a model reads, and warns that rough conversions such as dividing character count by four are inaccurate for real prompts. A token is closer to a fragment of a word than to a whole word, and the split depends on the specific text, not a rule a reader can eyeball.

What the documents say

Anthropic's token counting documentation states that Claude 4.7 and later models use a newer tokenizer producing roughly 30 percent more tokens for the same input, and warns against reusing a count measured on one model to judge cost or context fit on another. Google's token documentation gives a rough figure for Gemini: about four characters per token, or 60 to 80 English words per 100 tokens, and states images, audio and video convert to tokens too, with a small image counted as 258 tokens and video at roughly 100 to 300 tokens per second. No vendor publishes one conversion rate holding across languages, code and formatting; each documents an endpoint for counting tokens before a request is sent.

Check this

Before trusting a rule of thumb, a reader can run the counting tool a vendor actually documents rather than guess. Pasting a paragraph of plain prose and a paragraph of code or a table into a token counter is a direct test: dense or unusual text typically produces more tokens per character, since the split favors common fragments seen often in training. Anyone who switched models recently and found a prompt suddenly failing to fit should recount it, since the same words are not guaranteed to produce the same token total on a new tokenizer.

What holds and what fails

The instruction to count rather than guess holds whenever a workflow is tight against a context or billing limit, and it fails only in the trivial sense that most short chats never approach a limit worth measuring. It is editorial, not vendor guidance, to say a reader pasting a long document should treat the token, not the word, as the true budget: the documents establish the mechanism, not the advice to check it routinely. Reused token counts across model versions are the clearest failure mode the documentation flags, since a migration can silently change how far a prompt reaches.

• Run a vendor's own token-counting tool on a real prompt before assuming it fits.
• Recount prompts after switching model versions rather than reusing an old estimate.
• Treat images, audio and video in a prompt as token costs too, not free attachments.

The token is a modest piece of vocabulary, but it is the unit every limit, price and truncation in these systems is built from. Reading the documentation that defines it turns a vague sense that the model ran out of room into a specific, checkable number.