Token

What it means

When you type a question to an AI system, it doesn't read your words the way you do. First it breaks the text into small pieces called tokens. Anthropic, the company that makes Claude, says tokens "can correspond to words, subwords, characters, or even bytes."¹ NVIDIA gives an everyday example: the word darkness "would be split into two tokens, 'dark' and 'ness.'"²

The word matters to executives for one reason: tokens are how AI gets measured and billed. Google says "the cost of a call to the Gemini API is determined in part by the number of input and output tokens."³ (An API is the connection your company's software uses to send work to an AI model.) Google's rule of thumb is that "100 tokens is equal to about 60-80 English words."³ Your question counts as tokens, and so does the AI's answer. An AI agent that reads documents, checks its own work, and tries again uses tokens at every step.

Where the word comes from

The word is older than computers. In philosophy, it comes from a distinction the American philosopher C.S. Peirce drew in 1906 between a type and a token.⁴ The Stanford Encyclopedia of Philosophy illustrates it with a note that says "Remember: homework, homework, homework!" and explains: "When we count word types, the answer is two. When we count word tokens, the answer is four."⁴ A token is one instance of something, and each one is counted.

Computer scientists who work with language took up the word for the units their programs count. Its modern meaning, a piece of a word, owes a great deal to a 2015 paper by Rico Sennrich, Barry Haddow, and Alexandra Birch. They described a way of "encoding rare and unknown words as sequences of subword units," based on "the byte pair encoding compression algorithm."⁵ Breaking words into pieces lets a model handle names and words it has never seen. Today the same small unit is also what AI companies count when they bill customers.³

How others define it

• Anthropic: "Tokens are the smallest individual units of a language model, and can correspond to words, subwords, characters, or even bytes (in the case of Unicode)."¹
• Google: "Gemini and other generative AI models process input and output at a granularity called a token."³
• NVIDIA: "AI tokens are tiny units of data that come from breaking down bigger chunks of information."²
• Stanford Encyclopedia of Philosophy (the older, philosophical sense): it explains the type-token distinction by counting the words in a note. There are two types and four tokens.⁴

Where they agree and differ

They agree that a token is a small unit that gets counted. They differ on how big the unit is, and that difference affects your bill. Google says "a token is equivalent to about 4 characters" for its Gemini models.³ Anthropic says that on its earlier models "a token represents approximately 3.5 English characters," and that its newer models produce "approximately 30 percent more tokens for the same text."¹ So the same page of text can count as a different number of tokens at different companies, and even on different models from the same company. Comparing price per token across vendors can therefore mislead you.

They also differ on what tokens stand for. NVIDIA, which sells the chips that produce tokens, calls them the "currency" of AI and says AI factories turn them into "intelligence."² That's the strongest case for treating tokens as a measure of value: more tokens, more work done. The fair reply is that a token measures volume, not quality. An agent that loops ten times on a task it can't finish uses a lot of tokens and delivers nothing.

In the news

On October 26, 2026, the data company Cloudera released a survey of 1,500 IT leaders, run by Wakefield Research. It found that "nearly nine in ten (89%) respondents have already limited or redesigned AI use cases because of rising token consumption."⁶ Reporting on the survey, TechCentral wrote that "only 19% of them continuously monitor what they spend on tokens and inference" (inference means running an AI model to get answers).⁷ How sure we are: this is a survey commissioned by a company that sells AI infrastructure, so treat it as a strong signal, not settled fact. We found the 19% figure in TechCentral's report but not in Cloudera's press release.

What it means for your organization

Tokens turn AI from a fixed software license into a meter that keeps running. The people affected first are the teams whose useful projects get cut back when the bill arrives, and the finance staff who have to explain a cost nobody tracked. Price per token is only half of the bill. The other half is how many tokens your work uses, and AI agents that run many steps use far more than a single question to a chatbot.

One question to ask your team: For our most-used AI tool, do we know how many tokens one finished task uses, and who checks that number each month?