What are tokens in AI?
Tokens are the small units of text, such as words or parts of words, that an AI model reads and writes. AI usage and pricing are usually counted in tokens.
Why it matters
A tokeniser splits text, code or other input into subword units. Common words stay whole and rarer words break into pieces. Each token maps to a number the model can process. Input tokens are read in the prefill phase, and output tokens are generated one by one in the decode phase.
Providers bill AI usage by the token. Token counts by model and token type are also the starting point for working out the energy, carbon and water of AI use.
An example
Generating an output token takes more than ten times the energy of processing an input token. Inputs are used in such large volumes that the two often come out close in total. In Greenpixie's October 2026 data, a chat request to a large frontier model with 1,000 input and 500 output tokens uses about 1.9 Wh, including data centre overhead and a share of training energy.
How Greenpixie measures it
Greenpixie's AI methodology uses kWh per token as its main unit of energy. Each model's figures are matched to the provider's billing lines on three points:
- token type, covering input, output, cached input and embedding tokens
- processing tier, covering real-time, batch, priority and flex
- whether prefix caching applies, since a cached prefix is not processed again
Carbon and water per token are then worked out from the energy figure for each region, each given as a median with an interquartile range.
Sources
- Hugging Face Transformers: Tokenization algorithms
- Greenpixie AI methodology: 3. Definitions
- Greenpixie AI methodology: 2. Scope and Applicability
- Greenpixie AI methodology: 11. Outputs and Interpretation
- Greenpixie AI energy guide: How much energy does AI use?
- Greenpixie AI methodology: 7. Assessment and Calculation