What a token is and why it sets your AI bill

A taxi meter does not charge by the trip. It charges by the distance and the minutes, in small steps you can watch tick over. Per-use AI pricing works the same way, and its small step is called a token.

If you buy AI by use, almost every price you see is quoted per token, usually per thousand or per million. This piece explains what a token is, how a request becomes a count of tokens, and why that count is what sets your bill.

A token is a piece of text

A language model does not read whole words the way people do. Before it sees your text, a tool called a tokenizer cuts it into small pieces. Each piece is a token.

A short, common word is often a single token. A long or unusual word may be split into several. Spaces, punctuation, numbers and code all become tokens too.

The model then works one token at a time. It reads every token you send, and it writes its answer one token after another. That is why tokens are a natural unit for measuring the work.

Why the same text can be a different count

Each model family has its own tokenizer, so the same sentence can produce a different count from one model to the next. Languages differ as well. Text in some languages tends to break into more tokens than the same meaning in English.

Format matters. Tables, code and structured data can use more tokens than plain sentences of similar length.

Hidden text counts. Instructions an application adds behind the scenes, such as a standing system prompt, are tokens you pay for even if you never see them.

History counts. In a chat, earlier messages are often sent again with each new turn. A long conversation can cost more per reply than a short one.

From tokens to a bill: an illustrative example

The numbers below are made up and rounded to keep the arithmetic easy. They are illustrative only and are not any provider's real prices.

Imagine a service that charges 1 credit for every 1,000 tokens it reads and 4 credits for every 1,000 tokens it writes. You send a request of 2,000 tokens, including a document to summarise. The model replies with 500 tokens.

• Reading: 2,000 tokens at 1 credit per 1,000 = 2 credits
• Writing: 500 tokens at 4 credits per 1,000 = 2 credits
• Total for this request: 4 credits

Now run that request 10,000 times a month. The bill is 40,000 credits. Nothing about the single request looked expensive, but volume turned small steps into a real line item.

Notice that reading and writing were priced differently in this example. That is common in practice, and our piece on input tokens and output tokens explains why.

What drives your token count

A few choices move the count more than people expect. Most of them are within your control.

Prompt length. Long instructions and pasted documents raise the input count on every call.

Answer length. Asking for a short answer, or setting a limit, caps the output count.

Retries and chains. Some applications call a model several times for one user action. Each call has its own count.

Model choice. Different models charge different rates per token, and may use different numbers of tokens for the same text. See Cost per million tokens: how to read an AI price list.

Questions to ask before you compare prices

A per-token price is only useful if you know what it covers. Before comparing two offers, it helps to ask:

1. Is the price per thousand tokens or per million, and is it the same unit on both offers?
2. Are input and output tokens priced separately, and at what rates?
3. Which tokenizer is used, and how many tokens does my typical request produce with it?
4. Are system prompts, chat history or retrieved documents counted as input?
5. What date does this price apply from, and when was it last checked?

The last question matters more than it looks. Token prices change often, so a price without a date can mislead. We explain why in A price without a date is a rumour: why we publish dated prices. For dated prices and speeds across models, see Model Watch.

Tokens are not the whole cost

Tokens measure the model's work on your text. They do not capture everything you pay for. Network transfer, storage, monitoring and the people who run the system can sit outside the per-token rate. We cover these in Hidden costs of AI capacity that aren't on the quote.

Still, for most per-use buyers, tokens are the main moving part. Measuring a typical request, then multiplying by expected volume, is a sound first estimate.

What we are building

We think buyers should be able to see what they used and what it cost, in units they understand.

We're building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference. Workloads are designed to be placed by policy, on latency, data residency, model and load, and the network is designed so every workload can generate an auditable record of commitment, placement and delivery.

A clear record of the work done is the first step toward a bill you can check.

If you are planning AI inference and want to understand what it will use, use Request capacity below.