What a Token Actually Is
Language models don’t read words or characters — they read tokens, subword chunks produced by byte-pair encoding. Common words are one token; rarer words split into several; spaces attach to the word that follows; emoji and non-Latin scripts can take multiple tokens each. That’s why the same 500-word email can be 600 tokens of plain prose or 900 tokens of dense code, and why counting beats any words-times-a-constant rule.
Exact Where Possible, Honest Where Not
OpenAI publishes its tokenizer vocabularies, so GPT counts here are computed with the genuine article and match API billing for the text itself. Anthropic and Google don’t publish theirs, so no browser tool can count Claude or Gemini exactly — this one labels those figures as estimates (typically within ±15%) and the dedicated Claude and Gemini pages explain how to get exact numbers from each provider’s API.
Tokens Are Money and Memory
Two practical things are denominated in tokens: cost, since API pricing is per million tokens in and out, and memory, since a model’s context window caps how much it can consider at once. In multi-turn chats the entire history is re-sent on every turn, so conversations grow quadratically in billed tokens — the counter makes that growth visible before your invoice does. For turning counts into dollars across providers, the AI API cost calculator picks up where this page leaves off.
