What is an AI token, and how do token counts affect cost?

A token is the small unit an AI model actually reads and writes. It is not reliably one word or one character. The count depends on the model’s tokenizer, the rule set that breaks text into usable pieces. That count determines two practical things: how much material fits in one request and what an API user pays for it.

A token is a piece of text, not a word

People write in words, computers store characters, and language models work with tokens. A token can be a whole short word, a fragment such as the ending of a longer word, punctuation, or a space joined to nearby text. This lets a model handle familiar patterns efficiently without needing a separate entry for every possible word.

Google’s current Gemini token guide gives useful English approximations: one token is about four characters, and the word count is usually between three fifths and four fifths of the token count. Those are planning figures, not universal conversions. A number, an unusual name, a line of code and a sentence in another language can split differently. Models from different generations can count the same paragraph differently because their tokenizers differ.

That is why a word counter cannot settle a token budget. It can tell you that a report is long, but it cannot tell you the exact bill or whether a prompt fits a model. For exact work, count with the tools supplied for the model you will actually call.

Where the token count comes from

A tokenizer runs before the model. It maps pieces of your request to numeric identifiers the model was trained to process. Tokenization is that conversion from readable material into identifiers. The model then predicts an output token, adds it to the sequence and predicts the next one. The text you see is assembled from those output pieces.

The visible prompt is not always the whole input. OpenAI’s token counting guide distinguishes the text you type from other request material, including instructions and tool definitions. Anthropic’s counting endpoint likewise estimates tokens for system instructions, messages, tools, images and PDF documents. A system instruction is a standing rule supplied to the model before your request. A tool definition tells a model how it may call another service, such as search or a calculator. Both occupy part of the input even when the app does not display them beside your words.

Counts can also be estimates before a request. Anthropic says its counting endpoint is intended for planning and may not exactly match the final billed input. That distinction matters near a hard limit. Leave some room rather than designing a request that fits only if the estimate is exact.

Tokens set both capacity and price

The first limit is the context window, the maximum token sequence a model can consider in one request. It includes the conversation, attached material, hidden instructions, tool results and the reply. Our guide to context windows explains why a long chat can lose its beginning even when the final question is short.

The second limit is money. API price pages quote dollars per million tokens and usually separate the meter into three lines. Input tokens are what the model receives. Output tokens are what it generates. Cached input tokens are repeated input the vendor can reuse at a lower published rate. Our guide to AI model pricing takes those lines apart, while the pricing comparison shows the current rates stored in our ledger.

The chart above uses that ledger to show the cheapest tracked input rate at each vendor. It should not be read as a quality ranking. A low rate identifies a cheap meter among the models we track, not the best model for a task. Output rates, context surcharges, cache terms and tool charges can change the final cost. Compare the full row before choosing a model.

Consumer subscriptions hide most of this arithmetic. ChatGPT, Claude and Gemini normally express access through plan and usage limits rather than a token invoice. The underlying models still process tokens. The product simply translates that resource into a simpler allowance. The plan comparison records what each vendor publishes without converting unpublished limits into invented token counts.

How to estimate tokens before you send

For a quick English estimate, multiply words by about 1.3, or divide characters by about four. The token count for an ordinary English document will often be about one third higher than its word count. Treat the result as a range. Tables, source code, URLs, numbers and languages other than English can move it enough to matter.

Use an official counter when the request is large, repeated or expensive. OpenAI documents a token counting method for supported API inputs. Anthropic provides a token-counting endpoint before a message is created. Google reports counts through its Gemini token tools and returns usage metadata with responses. Usage metadata is the machine-readable count attached to a completed request. Logging it is more reliable than estimating from the visible answer after the fact.

A small worked example shows why both sides matter. Suppose a request contains 20,000 input tokens and asks for a 1,000-token answer. The input cost is 0.02 times the published input rate per million; the output cost is 0.001 times the output rate. Do those two calculations separately. Output is often priced higher, so asking for a concise answer can save more than trimming a few lines from a long source document.

What it means for you

If you use a chat product, do not count every casual message. Start a new conversation when the topic changes, avoid repeatedly pasting material the product already has, and summarise a long working chat before it drifts. Those habits reduce the context carried into later turns without turning normal writing into shorthand.

If you use an API, record input, cached input and output separately. Set a budget in dollars, translate it into a token range using the actual model rates, and watch the returned usage metadata. Repeated instructions should sit in a stable prefix where the vendor’s caching rules can recognize them. A prefix is the unchanged beginning of a request.

Most importantly, compare published figures rather than rules of thumb. The vendor hubs for OpenAI, Anthropic and Google connect their models, plans and status pages. Their tokenizers and prices can change. A dated source and an exact counter tell you what applies now; the idea that a token always equals a word does not.