What Is a Token in AI, and Why You Pay Per Token
Tokens are the units AI models read and write, and the units you get billed for. Here is what a token is, why output costs more than input, and how chats quietly add up.
Ask what is a token in AI and you usually get a circular answer: a token is the unit of text a model processes. True, but useless if you are staring at an API bill or a usage meter. Here is the version that actually helps. A token is a chunk of text, roughly three quarters of a word, and every AI chatbot charges you by the token because that is the only thing the model really counts.
Once you see text the way a model sees it, the billing stops feeling random. You start noticing why a long answer costs more than a long question, and why a chat that has been running all afternoon is quietly more expensive than the one you just opened.
What is a token in AI, really
A model does not read letters or words the way you do. Before any text reaches the model, a tokenizer breaks it into tokens. A token can be a whole word, a fragment of a word, a single character, or a piece of punctuation. Each unique token in the model's vocabulary gets an integer ID, so your sentence becomes a sequence of numbers the model can do math on.
GPT models use a method called Byte-Pair Encoding, which splits text into common subword pieces. The word "tokenization" might become two or three tokens. "Cat" is probably one. A rare word or a typo can get chopped into several small pieces, which is part of how models cope with text they have never seen before.
The handy rule of thumb for English is that one token is about four characters, or roughly 0.75 words. So 100 tokens lands around 75 words. Treat that as an approximation, not a law. The ratio shifts with the language, and code or rare words tokenize less efficiently than plain English prose.
Input vs output tokens
Every request splits into two buckets, and AI providers price them separately. Input tokens are everything you send: your prompt, the system instructions, any documents, and the running conversation history. Output tokens are everything the model writes back.
The output side works one token at a time. The model predicts the next token, appends it, feeds the longer sequence back into itself, and predicts again. That loop runs until the answer is done. It is the reason a long, detailed response feels slower and ends up costing more than a short one.
Input tokens: your message, the system prompt, uploaded files, and the chat history sent on every turn.
Output tokens: the model's reply, generated one token at a time.
Both are counted, billed, and capped by rate limits measured in tokens per minute.
Token pricing, and why output costs more
Providers quote prices per million tokens, with output almost always priced higher than input. In the table below, output runs five to six times the input price as of September 2026, and the exact multiple varies by vendor and changes often, so read these as point-in-time snapshots rather than fixed truth.
Model | Input per 1M | Output per 1M |
|---|---|---|
GPT-5.5 | $5.00 | $30.00 |
Claude Opus 4.8 | $5.00 | $25.00 |
Claude Sonnet 5 | $3.00 | $15.00 |
Those are standard API rates as of September 2026 (GPT-5.5 since April 24, Claude Opus 4.8 since May 28, and Claude Sonnet 5 after its introductory $2 and $10 rate ended on August 31, 2026). The shape is what matters: output is a few times pricier than input on every one of them. Why? Generation is the expensive part. Reading your prompt is a single pass over fixed text, while writing a reply means running the model again for every single token it produces.
This is also why controlling output length is the single most effective cost lever you have. Asking for a tight answer instead of an essay can cut the priciest part of the bill in half. If you want the deeper version of that, we wrote a whole guide on the cheapest way to run GPT-5, Claude, and Gemini, plus a breakdown of the cost to use all AI models side by side.
How context and memory quietly eat tokens
Here is the part that surprises people. The model has no memory of your conversation between turns. To keep the thread coherent, the entire chat history gets re-sent as input tokens on every single message. Turn one is cheap. Turn forty is dragging the whole transcript along for the ride, and you pay input rates on all of it, every time.
All of this lives inside the model's context window, the fixed token budget that input and output share. As that history grows, two things happen: each turn costs more, and quality can actually slip because models handle very long inputs less reliably. Stored memories and pinned instructions count too, since they ride along as input on relevant turns.
There is a practical fix. In MultiChats, when a thread gets unwieldy, a paid-plan helper flags that the chat is getting long and offers to summarize and start fresh, which trims the token load you carry forward. You can also switch models mid-conversation, so a lighter model handles the back-and-forth and a frontier model steps in only when you need the heavy answer. For the pricing details behind those model tiers, see our notes on cheapest GPT-5 and Claude access.
Frequently asked questions
How many words is one token?
Roughly 0.75 words in English, or about four characters. So 100 tokens is around 75 words, and 1,000 tokens is around 750. It is an approximation. Code, punctuation-heavy text, rare words, and non-English languages all tokenize differently, often using more tokens for the same visible length.
What is the difference between input and output tokens?
Input tokens are everything fed into the model: your prompt, system instructions, files, and the prior conversation. Output tokens are what the model writes back. Providers count and bill the two separately, and they almost always charge more for output.
Why is output more expensive than input?
Because generating text is harder work than reading it. Input is processed in one pass. Output is built one token at a time, with the model running again for each new token. Providers price output several times higher than input: for the models in this article's table, five to six times as of September 2026, with the exact multiple varying by vendor and changing often.
Do longer chats cost more?
Yes. Since the full history is re-sent as input on every turn, a long conversation costs more per message than a fresh one, even before the model replies. Summarizing and starting a new thread, or trimming what you keep in context, brings that creeping cost back down.
The short take
Tokens are just text cut into small pieces, priced per million, with output costing more than input and long chats quietly stacking up input tokens turn after turn. Understand that, and the meters stop feeling like a black box. MultiChats puts 25+ models behind one flat subscription, so you can run lighter models for the chatter and frontier models for the hard parts without juggling separate per-token bills. Compare the MultiChats plans and pricing and skip the per-token math.