AI Tokens
AI tokens are the units into which a language model splits text — whole words, parts of words, punctuation or bytes — before processing it, and the unit in which model use is measured and billed. Providers price input and output tokens separately, usually per million, so token volume drives the running cost of any AI product.
- Also known as
- LLM tokens, Language model tokens, Tokens (AI)
- Size
- about four characters of English, with 100 tokens equal to roughly 60–80 English words, by Google's guide for Gemini; Anthropic puts a Claude token at about 3.5 English characters; the ratio varies by language and model
- Pricing
- quoted per million tokens, with output usually several times dearer than input, and discounts for cached input and for batch processing
- Limits
- a model's context window (how much it can read at once) and the maximum length of its reply are both set in tokens
- Hidden tokens
- system instructions, tool definitions and a reasoning model's internal "thinking" all count; Anthropic, for example, bills thinking as output tokens
- Not to be confused with
- payment tokens, which stand in for card numbers, or crypto-asset tokens on a blockchain
How AI tokens work in practice
Every call to a model is metered. The prompt — the user's question plus system instructions, retrieved documents, conversation history and tool definitions — is counted as input; the reply, including any reasoning the model does before answering, is counted as output. Multiply each by its price and the cost of one interaction appears; multiply by volume and it becomes a line in the cost of sales.
That arithmetic is why token costs matter to anyone valuing an AI-heavy business. A support assistant that pastes 20 pages of policy into every prompt pays for those pages on every message; one that retrieves the two relevant paragraphs pays for two. Long conversations grow more expensive as history accumulates. Reasoning models may spend thousands of output tokens thinking before a short answer. Engineering teams manage this with prompt caching, retrieval, smaller models for easy tasks and caps on output length, and finance teams track cost per resolved ticket or per document rather than per call.
For a buyer, the questions are what share of revenue goes to model providers, how that share moves as usage grows, how exposed the margin is to a provider's price change, and whether the product's pricing — per seat, per task or per outcome — passes token costs through or absorbs them. A falling price per token does not guarantee a falling bill if each task starts using more of them.
Example
A lending platform uses a model to summarise bank statements for its underwriters. Each application sends 30,000 input tokens (statements plus instructions) and gets back a 1,500-token summary. At $3 per million input tokens and $15 per million output, each summary costs $0.09 plus $0.0225 — about $0.11, or $4,500 a month across 40,000 applications. The team caches the fixed 6,000-token instruction block, charged at a tenth of the input price on reuse, and trims statements to the relevant months, cutting input to 18,000 tokens; cost falls to about $0.06. Then it switches on a reasoning mode that adds 4,000 thinking tokens per case. Those are billed as output, adding $0.06 each, and the bill ends above where it started.
Common mistakes
- Estimating cost from the user's question alone and forgetting the instructions, history and documents sent with it. - Assuming a token is a word; in English 100 tokens are roughly 60–80 words, and other languages and code differ. - Ignoring reasoning tokens, which can dwarf the visible answer. - Comparing providers on input price when output dominates the bill. - Pricing an AI product per seat with no usage limit, so the heaviest customers become unprofitable. - Mixing up AI tokens with payment or crypto tokens in a deal document.