What a token actually is
AI language models don't read letter by letter or word by word. They break text into chunks called tokens, which can be a whole word, part of a word, or a punctuation mark. In my token comparison in Choosing Your First AI Tool, all three assistants gave the same rule of thumb: one token is about 4 characters of English, or roughly three-quarters of a word.
Turning that into numbers you can use
- 1,000 words is roughly 1,300 tokens.
- A 3,000-word article is roughly 4,000 tokens.
- A 300-page book is on the order of a hundred thousand tokens or more, depending on the text.
These are estimates. For an exact count, paste your text into the tokenizer tool that many AI providers offer on their websites.
Not all languages cost the same. Text in languages such as Hindi or Marathi often needs more tokens per word than English, because the tokenizer was built mostly around English. The same content can therefore use up more of your limit, or cost more, in some languages.
What a "context window" is
The context window is the maximum number of tokens the model can consider at once. Think of it as the model's desk: everything it can see has to fit on it. That includes your whole conversation so far, any documents or images you shared, hidden instructions from the app, and the answer it is writing right now.
Once the desk is full, something has to come off. Depending on the tool, older messages may be dropped or summarized, and the model genuinely cannot see them anymore.
Context window is not memory
These are easy to mix up. The context window is what the model can see in this conversation. A memory feature is something a tool adds on top, saving selected details and putting them back into a new conversation's context. A model with a huge window and no memory forgets you the moment you start a new chat.
Why context windows have grown so much
Early language models handled a few thousand tokens. Modern models often support hundreds of thousands, and some support over a million. That lets you work with an entire report, book or codebase in one conversation.
Practical effects of hitting the limit
- The model "forgets" early parts of a long chat. It isn't being careless. That text is no longer on the desk.
- Pricing is often based on tokens, for both what you send and what comes back. Long documents and long chats cost more per reply.
- Quality can slip in very long inputs, even inside the stated limit. Details buried in the middle are the easiest to miss.
Habits for long tasks
- Put the most important instruction at the start and repeat it near the end.
- For a very long chat, ask for a summary of the key decisions, then start a fresh conversation with that summary.
- Send only the parts of a document that matter, not the whole thing, when you can.
Frequently asked
How do I know how many tokens my text uses?
Most AI provider websites offer a "tokenizer" tool where you can paste text and see the exact token count — useful before working with anything close to a context limit.
Do images and files count as tokens too?
Yes, though the conversion varies by tool — images and documents typically get converted into a token-equivalent count that also uses up your context window.
Is a bigger context window always better?
It's more capability, but not automatically better performance — some models handle very long contexts less reliably than their stated maximum might suggest, especially for details buried in the middle of a long input.