What Is a Token in AI, in Plain Words: Cost, Context Window, How to Estimate
AI tokens in plain words: what they are, why you pay for them, how input differs from output, what a context window and limits are, and how to estimate the cost of your task.
AI models read and write not in letters or words but in tokens. Tokens are what you pay for, they measure how much text a model can keep in mind, and every limit is set in them. Once you understand tokens, it becomes clear why one request costs a fraction of a cent and another costs dollars, and why a chatbot suddenly "forgets" the start of a conversation.
A token in plain words
A token is a piece of text the model works with. Sometimes a whole short word, sometimes part of a long word, sometimes punctuation or a space. Before processing, text is cut into these pieces, and that is what the model sees.
Think of building blocks: the model cannot handle text as a whole, it assembles and takes it apart from ready-made pieces. Frequent words and combinations are large pieces; rare words are built from several small ones.
How AI splits text into tokens
Each model has its own vocabulary of pieces, so the same text takes a different number of tokens in different models. General patterns:
- frequent short words usually take one token;
- long and rare words are split into several parts;
- digits, punctuation and spaces also take tokens;
- for English, a common rule of thumb is one token per 3-4 characters;
- some languages, Russian among them, are usually cut into smaller pieces, so the same meaning tends to take more tokens.
Practical takeaway: do not guess the token count; measure it on a few real examples of your text.
How text becomes an answer
- 01Your textprompt and data
- 02Splittingtext is cut into tokens
- 03Modelreads the input tokens
- 04Answerwrites tokens one by one
- 05Billinginput and output counted separately
Input and output: what you pay for
| What it is | Price | |
|---|---|---|
| Input | your prompt, instructions, documents, chat history | cheaper |
| Output | what the model writes | usually 3-5 times the input price |
| Cached input | a repeated prompt prefix the model has seen | noticeably cheaper than regular input |
Prices are quoted per 1 million tokens. From official price lists as of 28 October 2026:
| Model | Input per 1M | Output per 1M |
|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 |
| DeepSeek V4.1 Flash | $0.30 | $1.20 |
| GPT-6.1 Terra | $1.00 | $5.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
DeepSeek charges half price off-peak; the Claude Haiku price applies to prompts up to 100K tokens. More models and prices are in the article on choosing an AI model.
What a token costs: two estimates
Channel posts. Each post takes about 1,500 input tokens (instructions plus source news) and 500 output tokens.
| Model | 1 post | 1,000 posts |
|---|---|---|
| DeepSeek V4.1 Flash | about $0.001 | about $1 |
| Claude Opus 5.5 | about $0.016 | about $16 |
A support chatbot. 1,000 conversations a day, each answer takes about 3,000 input tokens (instructions, knowledge base, history) and 300 output tokens. That is 30,000 answers a month.
| Model | Per month |
|---|---|
| DeepSeek V4.1 Flash | about $38 |
| GPT-6.1 Terra | about $135 |
| Claude Sonnet 5.5 | about $270 |
The difference between models reaches tens of times. That is why production systems use inexpensive models for simple bulk tasks and keep expensive ones for where they are really needed.
How to estimate tokens for your task
- Take 5-10 real requests: instructions, data, expected answer.
- Count tokens with the provider's counter or from usage after test requests.
- Multiply the average by the number of requests per month.
- Calculate input and output separately at your model's prices.
- Add a 20-30% buffer for long requests, retries and growth.
What a context window is
The context window is how many tokens a model can take into account at once: instructions, documents, chat history and its own answer together. Modern windows can be very large, for example about 1 million tokens for DeepSeek V4 and Kimi K3.
What happens when the text does not fit:
- the request is rejected;
- in chats, old messages are dropped and the bot "forgets" the beginning;
- in long documents, the model notices details in the middle less well.
A large window is convenient, but every token in it is billed. Instead of putting the whole archive into the prompt, find and pass only the relevant pieces. That is how RAG works; it has its own article.
Token limits
- Maximum per answer. Models cap the length of a single answer. A long report sometimes has to be assembled from several parts.
- Per-minute limits. Providers limit tokens and requests per minute per account. Plan for them in bulk tasks.
- App limits. Subscription chat apps count messages or "usage" rather than tokens, and heavy work can exhaust them.
Hidden tokens people forget
- Reasoning. Models that "think" before answering spend tokens on it, billed as output.
- Chat history. In a chat, every new message resends the whole previous conversation.
- Instructions and tools. An agent's system instruction and tool descriptions go into the input of every request.
- Images, audio and video. In multimodal models they are also converted into tokens and billed, often at separate prices.
How to save tokens
- Give the model only the relevant part of a document, not the whole archive.
- Put the standing instruction at the start of the prompt so caching kicks in.
- Ask for a short answer in a set format, not an essay.
- Trim or summarise old chat history.
- Use cheap models for simple tasks and expensive ones for hard tasks.
- Send bulk work that does not need an instant answer through batch processing: many providers charge half for it.
- Track spend daily and set limits so an automation bug cannot burn the budget overnight.
FAQ
What is a token in AI in simple terms?
A piece of text into which the model splits the prompt and the answer. It is also the unit for measuring text volume and cost.
How many words are in a token?
There is no exact answer: it depends on language and model. For English, about 3-4 characters per token is a common estimate.
Why is output more expensive than input?
The model generates the answer one token at a time, which takes more computation than reading the prompt.
What is a context window?
The maximum amount of text in tokens a model considers at once, including instructions, documents, history and the answer.
Why does a chatbot forget the start of a conversation?
The whole conversation no longer fits in the context window and old messages are dropped. Summarising history and storing key facts separately helps.
How do I know how many tokens a text has?
Many providers offer token counters in their docs or APIs, and actual usage is visible after requests. For budgeting, estimates on a few real examples are enough.
Are ChatGPT tokens and API tokens the same?
The unit is the same, but billing differs: the app is a subscription with usage limits, the API charges per million tokens.