What Is a Token? The Essential AI Primer You Must Know

In March 2026, the National Data Administration officially defined a token in the AI field as a “词元” (semantic unit). A token is the smallest semantic unit that large language models handle; the model does not process raw text directly but first splits the input into a token sequence for computation.

The complete processing flow is: first, segment the text into tokens; then map each token to a corresponding numeric ID; finally, convert the IDs into vectors that the model can recognize. Only after these three steps does the model begin to process the content.

Unlike humans who read a whole sentence at once, AI processes text more like a series of small blocks. For example, the sentence “今天天气真不错” is a complete phrase for a person, but the AI first breaks it into several small pieces—these pieces are the tokens.

Analogies help illustrate the concept: just as a whole loaf of bread is inconvenient to eat in one bite, a whole paragraph is inconvenient for the model, so it is sliced into tokens. Each slice corresponds to a token, which the AI then processes piece by piece before reassembling the meaning of the entire sentence.

A token does not correspond directly to a character or a word. Roughly, one token equals about four English characters or three‑quarters of an English word, but this value varies with language and the tokenizer used. In Chinese, a single character can sometimes be a token, while two or three characters may also be combined into one token, depending on the tokenizer.

Understanding tokens is important because token counts determine two key aspects of AI usage. First, tokens define the context window size: the total number of tokens the model can consider in a single interaction, which directly limits how much content the model can remember at once. Second, tokens are the basis for billing: AI services charge per token, similar to how phone calls are billed per minute or shipments per weight. Both the input you provide and the output generated consume tokens.

These points only scratch the surface; the underlying logic of tokens is richer and more interesting than it appears at first glance.