AI Token

The unit on every AI bill. Model providers do not charge by word, page, or task — they charge by token, and almost every surprise invoice in enterprise AI is a token story.

What It Is

A token is the smallest chunk of text a language model reads or writes — sometimes a whole word, often a fragment. OpenAI's published estimate for English is that one token is approximately four characters, or about three-quarters of a word, so a 500-word page is roughly 670 tokens. Every category of text has its own line on the meter: input tokens (what you send), output tokens (what comes back), cached input (repeated prefixes, billed far cheaper), and, for reasoning models, internal thinking tokens you never see but pay for. API prices are quoted per million tokens, with output typically priced several times higher than input — at the time of writing, OpenAI's pricing page lists its current flagship at US$10 per million input and US$50 per million output tokens.

Why It Matters to You

Subscriptions hide the meter; APIs expose it. The moment a workflow calls a model directly — through an integration, an agent, or a vendor product that passes costs through — spend scales with volume, document size, and how much context gets re-sent on every step. Two rules of thumb save most executives: reading is cheaper than writing but happens far more often, and long documents multiplied by many calls dominate every bill.

When You'll Meet It

At the first API invoice, and in the name collision that catches even experienced buyers: this token has nothing to do with tokenization in the data-security sense — same word, unrelated concept. When a vendor quotes an AI price, ask: how many input, output, and cached tokens does one typical task consume, at which model's rates?