Key takeaways
- A token is a small piece of text — often part of a word. Models read, write and charge in tokens, not words.
- In English, one token is roughly 3–4 characters, or about three-quarters of a word. Arabic usually needs more tokens for the same meaning.
- API pricing is quoted per million tokens, and output tokens usually cost more than input tokens.
- The context window is how much text a model can consider at once — beyond it, earlier parts of a conversation are dropped.
What a token is
Before a language model can work with text, it splits it into tokens: common words may be a single token, while longer or rarer words are broken into several pieces. Spaces, punctuation and numbers count too. The model never sees “words” the way we do — only sequences of tokens, which it reads and then predicts one at a time (see our plain-words guide to AI for business for how that prediction works).
A rough rule for English: one token is about 3–4 characters, or roughly three-quarters of a word. So 100 tokens is somewhere around 75 words. The exact count depends on the model, because each provider uses its own tokenizer.
Why Arabic often costs more
Tokenizers are built mostly from the text they were trained on, and much of that is English. Arabic, and many other languages, are therefore often split into more, smaller pieces. The result: the same message written in Arabic usually uses more tokens than its English equivalent — which means it can cost more, run slightly slower and fill the context window sooner. Newer models have improved on this, but it is still worth knowing if your business works in both languages.
Input vs output tokens
Every request has two sides:
- Input tokens — everything you send: your question, any instructions, attached documents, and (in a chat) the earlier conversation that gets re-sent each turn.
- Output tokens — everything the model writes back.
Providers price these separately, typically in dollars per million tokens, and output tokens usually cost noticeably more than input. Prices differ widely between models and change often, so always check the provider's current pricing page rather than relying on a figure you read somewhere.
The context window
The context window is the maximum number of tokens a model can take into account at once — your instructions, the documents, the whole conversation so far, and its own reply. Modern models have large windows, but they are not unlimited.
This explains two things people notice in practice:
- Long chats get slower and costlier. In a chat, the earlier messages are sent back to the model with every new question, so each turn processes more input tokens than the last.
- Models “forget”. Once a conversation grows beyond the window, the oldest parts are dropped or compressed, and the model genuinely no longer sees them. Even within the window, details buried in a very long conversation can be overlooked.
| Term | What it means |
|---|---|
| Token | A small piece of text (often part of a word) that the model reads and writes |
| Tokenizer | The method a model uses to split text into tokens; it differs between providers |
| Input tokens | Everything sent to the model: prompt, instructions, documents, chat history |
| Output tokens | Everything the model generates in reply; usually priced higher |
| Context window | The maximum tokens the model can consider in one go |
| Prompt caching | A discount some providers offer when the same long input is reused across requests |
| Rate limit | A cap on how many tokens or requests you can use per minute or per day |
How to reduce token costs
- Keep prompts focused. Send the relevant section of a document, not the whole archive.
- Start fresh chats for new topics instead of continuing one endless conversation.
- Summarise as you go. For long projects, ask for a short summary and start a new chat from it.
- Ask for shorter answers when you don't need detail — output is the expensive side.
- Match the model to the task. Smaller, cheaper models handle classification, extraction and simple replies well; save the largest models for complex reasoning.
- Use caching where offered. If you send the same long instructions or reference document repeatedly, prompt caching can lower the cost of those repeated input tokens.
Subscriptions vs API usage
There are two common ways to pay. A per-user subscription to a chat app (such as ChatGPT or Claude plans) gives a fixed monthly price with usage limits — simple and predictable for staff using AI interactively. API usage is pay-as-you-go per token and is what you use when AI is built into your own software or automated workflows. Here, tokens become a direct line on your bill, so it is worth setting spending limits and monitoring usage from day one.
The bottom line
Tokens are the unit behind everything an AI model does: what it can read, what it remembers, how fast it answers and what it costs. Knowing roughly how they are counted lets you write better prompts, pick the right model and avoid surprise invoices.
If you're planning to build AI into your workflows and want help estimating costs and choosing the right setup, talk to our team.
