azyware
LLM application engineeringConcept

Token

Also: subword token

In one sentence

What is Token?

A token is the unit of text a language model reads and writes, typically a word fragment of a few characters, and the unit in which context limits and usage-based pricing are measured.

What Token means

A token is a chunk of text produced by the model's tokeniser, which splits input into pieces from a fixed vocabulary. Common English words are often one token; longer or rarer words, numbers, code and non-Latin scripts split into several. As a rough guide, one thousand English tokens is around 750 words, but Hindi, Kannada or Tamil text can cost two to four times as many tokens for the same meaning, depending on the model.

Tokens matter because everything is measured in them: the context window is a token limit, provider pricing is per million input and output tokens, and latency scales with tokens generated. Output tokens usually cost more than input tokens, and a verbose answer is a slow, expensive answer.

A token is not a word and not a character, and counts differ between models, so a prompt that fits one provider's limit may not fit another's. It is also not a unit of meaning: the model has no special understanding of token boundaries, which is why it can struggle with tasks like counting letters.

Who it really matters to

  • CFO: tokens are the billing unit for every LLM call, so forecasts and per-account cost reports are token forecasts.
  • CTO / Head of Engineering: token counts drive context design, caching decisions and model choice, and differ per provider.
  • Product manager: multilingual users cost more per interaction, which affects pricing and margin on AI features.
  • Operations head: response length and verbosity are controllable levers on both cost and speed.

Why it exists

Tokens exist because models cannot process raw characters efficiently and cannot know every word in every language; subword pieces are the compromise that keeps vocabularies small and coverage broad. For buyers, the term matters because it is the metering unit. Knowing that a long system prompt is re-sent on every call, that Indian-language text tokenises expensively, and that output costs more than input explains most surprises on the monthly bill. The trade-off is that token efficiency and answer quality can pull in opposite directions, so cost work has to be measured against evals.

Where it is applied

  • Forecasting the monthly bill for a SaaS copilot from expected requests, average prompt size and answer length.
  • Metering AI usage per tenant in a multi-tenant product so heavy accounts can be priced or capped.
  • Choosing a model for a Hindi and Kannada support agent based on tokens per message, not just list price.
  • Setting output limits on a logistics customer-update generator so every message stays short and cheap.
  • Estimating how many pages of a loan file fit in one extraction request.

Is Token a skill?

ConceptA concept rather than a skill, but one that underpins cost engineering. Eazyware's LLM applications work includes token budgeting, per-account cost reporting and model routing so token spend is a managed number rather than a surprise.

Eazyware service that covers it: LLM Application Development. Starting prices are on the pricing page.

Frequently asked questions

How many tokens is a typical page of text?

Roughly 500 to 700 tokens for a page of English prose, more for tables, code or dense formatting. Indian-language text on the same page can be two to four times higher depending on the model's tokeniser.

Why is the same prompt priced differently across providers?

Each provider has its own tokeniser and its own per-token rates for input and output. The same text can be a different token count on each, so comparisons should be made on measured cost per request, not list price.

Related reading

Need Token built, not just explained?

PRJECT IN MIND?