What Is a Token? Why AI Counts in "Tokens"

In the AI tools' documentation, price tables and error messages, always the same word: token. "A 200-thousand-token context window", "the price per million tokens", "the token limit exceeded". So what is a token? Understanding this measuring unit of the AI world is no technical curiosity but a practical need: the limits, the prices and every "why did it cut off midway" question hang on this concept.

This article explains the token in plain language: what it is, why the counting runs on it rather than on letters and words, what the context window means and which rules fall out for daily use.

The token: the model's "bite"

The model does not read text letter by letter or word by word as we do; it splits it into its own pieces, and those pieces are called tokens. A token is sometimes a whole word ("book"), sometimes a word fragment ("automa" + "tion"), sometimes a punctuation mark. Why so? A technical compromise: reading letter by letter creates very long sequences, while reading word by word demands storing millions of rare words separately; the fragment approach is the middle of the two: with a limited vocabulary you can assemble any text. The rough ratios (they vary by language): in English 1 token ≈ 4 letters / 0.75 words; in under-represented languages like Azerbaijani the same text usually eats more tokens (the words split finer); meaning slightly less text fits the same limit in our language. The LLM article's "predicting the next word" description sharpens: the model in truth predicts the next token.

The context window: the model's desk

The most important practical concept: the context window is the token volume the model can "see" at once; your messages + the documents you upload + the model's own answers must all fit that window. What happens when the window fills? The old parts turn invisible; that is why the model "forgets" the details from a long conversation's start: its memory has not broken — they left the window. In modern models the windows are large (hundreds of thousands of tokens: hundreds of pages), but not endless, and one nuance exists too: the information in the window's middle can be "recalled" weaker than at its edges. The practical rules: refresh the important context now and then in long conversations, put the question concretely next to the document when working with a large one, and for giant archives look at the RAG approach: instead of pouring everything into the window, finding and supplying what is needed.

The price and the limit: the token's money side

The conclusions for the business user: in API projects the cost forecast runs on the token count (the typical request × the monthly count × the tariff), while for the subscription user the rule is simple: do not carry needlessly long context; pasting the whole history into every request eats both the quality and the limit.

The daily-use rules

• Give the context selectively: not the whole document but the relevant part; half of a good prompt is the right context choice.
• Chunk the long work: processing a 50-page report section by section is both a quality and a limit friend.
• Refresh the conversation: opening a new chat when the topic changes drops the old load; the "everything in one chat" habit litters the window.
• Build the summary bridge: if a long conversation's continuation is needed, ask the model for a summary and carry it into a new chat; the professionals' standard trick.

Sources and further reading

Where to verify the source

The tokenisation tools:

• OpenAI Tokenizer: see your text token by token

Continuing the topic

The AI literacy line's neighbouring articles:

• The LLM: the base concept
• RAG: working with large archives
• The context and the prompt
• The AI cost calculation
• Other articles on this topic

To close with an analogy: the token is the model's bite, the context window its table. When you choose what you lay on the table (the relevant material, the clean context), the model "eats" well too; everything else derives from that simple rule.