THE SHORT ANSWER
A token is a unit produced by a model's tokenizer. It may be a word, part of a word, punctuation or another encoded element. Token counts vary with language, content and model. Providers often meter input and output separately, so count with the tokenizer or usage data for the model you actually use.
A token is not simply a word
Language models operate on encoded sequences. A tokenizer divides text into units from its learned vocabulary. A familiar word may be one unit; an unusual name may become several. Punctuation, spaces, code and non-English scripts can behave differently. A word-count estimate is therefore useful only for rough planning.
Different model families can use different tokenizers. When a limit or cost matters, use the provider's usage response or the matching tokenizer rather than a universal words-to-tokens conversion.
Evidence & context: Vaswani and colleagues, 2017 · Brown and colleagues, 2020
Separate input from output
| Category | Examples | Control |
|---|---|---|
| Input | Instructions, conversation, documents and tool results | Select relevant context and avoid repeated material |
| Output | Answer, structured data or generated code | Specify format, scope and an appropriate maximum |
| Hidden/provider-specific | Some systems report reasoning or cached-token categories | Read the provider's current usage definitions |
Two requests with the same question can use different totals if one carries a long conversation or produces a longer answer. That is why cost analysis needs the entire request payload and response, not the visible prompt alone.
Avoid three misleading shortcuts
- Tokens are not credits. Credits are a commercial entitlement defined by a service.
- Tokens are not requests. A single request can process very different amounts of information.
- Fewer tokens do not automatically mean a better workflow. Missing evidence can create retries and review work.
Inspect one real exchange
Take a repeated task and label every part of its input: stable instruction, current question, history, retrieved evidence and tool output. Then label the response as required or optional. This reveals which tokens carry task information and which arrived because the workflow never made a selection.
Sources & further reading
- Language Models are Few-Shot Learners
Brown and colleagues, 2020. Primary research on an autoregressive language model and learning from examples in context; not a guarantee about all models.
- Attention Is All You Need
Vaswani and colleagues, 2017. The original Transformer paper. Its experiments concern sequence transduction, particularly translation.
- Models
OpenAI Developers. Official model-selection documentation, checked 13 September 2026. Product names, capabilities and prices can change; the collection uses the durable principle of matching capability to a task rather than prescribing a current model.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗