The AI Handbook Open in the app →

Token

Inference & Prompting · Beginner · 4 min read

What is it?

A token is a chunk of text — often a word or part of a word — that language models read and generate one at a time.

Explain like I'm 5

Tokens are like the puzzle pieces of text: the model doesn't see whole sentences at once, it works piece by piece, where each piece is roughly a word or word-fragment.

Why was it created?

Models need text in consistent, manageable units. Tokenization was created to break language into pieces a model can process and predict.

Where is it used?

  • Measuring LLM input/output size
  • Pricing AI API usage
  • Enforcing context limits
  • Counting prompt length

Why should developers care?

Tokens determine model cost, speed, and limits, so anyone building with LLMs budgets and reasons in tokens.

How does it work?

A tokenizer splits text into tokens using learned rules, turning each into a number the model understands. The model predicts the next token repeatedly to generate text, then the tokens are turned back into readable words.

Real-world example

The word 'unhappiness' might be split into a few tokens like 'un', 'happi', 'ness', each processed separately by the model.

Common use cases

  • Estimating API cost
  • Staying within context limits
  • Measuring prompt size
  • Comparing model efficiency

Advantages

  • Consistent processing units
  • Handles unknown words via fragments
  • Enables clear cost/length measurement
  • Language-agnostic

Disadvantages

  • Token counts aren't intuitive
  • Differ between models
  • Some languages use more tokens per word
  • Easy to underestimate length

When should you use it?

Whenever you need to reason about LLM length, cost, or limits.

When should you avoid it?

Not something to avoid — it's an underlying unit, useful to understand rather than opt out of.

Alternatives

Counting characters or words (less accurate for models)

Related terms

Large Language ModelContext WindowTransformerPrompt Engineering

Interview questions

Beginner

  • What is a token?
  • Is a token always a whole word?

Intermediate

  • Why do models use tokens instead of words?
  • Why does token count affect cost?

Senior

  • How does tokenization affect non-English text cost?
  • Why might the same text differ in token count across models?

Common misconceptions

  • "One token equals one word" — a token is often a word fragment; long or rare words become several tokens.
  • "Token counts are the same everywhere" — different models tokenize differently.

Fun facts

  • A rough rule of thumb is that a token is about four characters of English text.
  • Both your prompt and the model's reply consume tokens.

Timeline

  • 2010s — Subword tokenization becomes standard for language models

Learning resources

Quick summary

A token is a chunk of text (often a word fragment) that models process one at a time, and the unit that governs LLM cost and limits.

Cheat sheet

  • Text chunk, ~a word or fragment
  • Models read/write token by token
  • Drives cost and context limits
  • Varies by model

If you remember only one thing

Models read and write text in tokens — word-sized chunks that determine cost and length limits.