$ learn
How does AI think?
Why does a reply from ChatGPT or Claude appear piece by piece? Because the model reads and writes in small units called Tokens, and Arabic splits into more of them.
The short version
When you ask ChatGPT or Claude for something, the reply doesn't appear all at once; it gets written in front of you bit by bit. That isn't a visual effect to dress up the interface. The model really is writing the reply while you watch. But it doesn't write it letter by letter, or word by word the way we do.
One term is enough to understand how the model reads what you say and how it writes its reply: Token. You will also see why Arabic text gets split into more Tokens than English, and why that is no reason to switch to English.
Key terms
- Token: a small unit of text (a word or part of a word) that the model processes text in. What you send counts as Input Tokens; what the model generates counts as Output Tokens.
- Tokenizer: a tool that shows how text is split into Tokens and how many there are.
The explanation
The keyboard that suggests the next word
When you type a message on your phone, the keyboard suggests what might come after what you have typed. You pick the suggestion, it suggests the next one, then the one after, until the sentence is complete.
The keyboard doesn't understand what you want to say. It only sees what has been typed and suggests what is likely to come next. The model works on the same logic, though it is far smarter and more complex: it picks the next piece, then the one after that, until the reply is complete.
What is a Token?
That small piece of text is called a Token. The model doesn't see words or letters the way we do; it sees Tokens. It reads what you say in them, and it writes its reply in them.
A Token doesn't have to be a whole word. Sometimes a single word is split into several Tokens, and sometimes a short word is a single Token.
Why does Arabic split into more Tokens?
These models were trained on far more English text than Arabic text. So common English words ended up with ready-made Tokens that cover them whole, while Arabic words get split into smaller, more numerous pieces.
The reason isn't that Arabic is harder, or that there is anything wrong with it. The model has simply seen more English.
In the video, we typed two sentences with the same meaning into a Tokenizer. The Arabic one is colloquial Egyptian for "AI writes the reply piece by piece". The result:
| Sentence | Characters | Tokens |
|---|---|---|
| الذكاء الاصطناعي بيكتب الرد حتة حتة | 35 | 16 |
| AI writes the reply piece by piece. | 35 | 8 |
Exactly the same amount of writing, and yet the Arabic sentence needed twice as many Tokens.
The reply is written Token by Token
In the second part of the video, we asked Claude to finish an incomplete sentence with a single sentence, and slowed the video down so you can follow the reply as it appears. The model doesn't prepare the whole reply and then show it to you; it puts it out one Token after another, just like the keyboard suggesting the next word.
What you send and what comes back
The Tokens you send to the model are called Input Tokens, and the ones it replies with are called Output Tokens. It is the same unit, once on its way to the model and once on its way back to you. You will come across these two terms a lot, because tools that charge by usage count each of them separately.
Try it yourself
Open OpenAI's Tokenizer at platform.openai.com/tokenizer. In the video we used the GPT-4o-and-later setting (the o200k_base encoding). Pick the same setting, because the numbers differ from one setting to another.
Paste the Arabic sentence below and look at the colours: each colour is a separate Token. Then look at the two counters: the number of Tokens and the number of characters.
الذكاء الاصطناعي بيكتب الرد حتة حتةClear it, paste the English sentence with the same meaning, and look at the two counters again.
AI writes the reply piece by piece.Compare the two results. In the video, both sentences had the same number of characters (35), while the Arabic came to 16 Tokens and the English to only 8.
Why this matters to you
Since the model reads and writes in Tokens, they are also the unit of measurement in tools that charge by usage. In those tools, three rules hold:
- Input Tokens and Output Tokens are counted separately.
- Output Tokens cost more than Input Tokens, often several times more.
- The counting method differs from tool to tool, so a lower price per unit can mean a higher bill if the tool splits the same text into more Tokens.
Chat subscriptions to ChatGPT and Claude, on the other hand, are flat-rate; you don't pay per Token. Counting in Tokens starts when you move to serious coding tools like Claude Code and Codex, or when you use the models through an API. Even there, your choice of language only starts to matter for cost with long instructions sent along with every request, not with ordinary chat messages.
Because prices change, we don't list them here. Always check them at the source:
Images you attach to a request are also counted in Tokens, but differently: by the size of the image, not the number of words.
Related lessons
- What is AI?: what Model, LLM and Chatbot mean.