Tokens: the unit of text
The model reads and writes not in letters or words but in tokens — chunks of text ranging from a single character to a whole word. "Hello" is one or two tokens; a rare long word is three or four. On average a token is about 4 characters for English text and about 2–4 characters for Russian.
Why a user should care:
- Limits are counted in tokens. The amount of text the model can read and write is measured in tokens, not pages.
- API pricing is in tokens too. Once we get to automation, you will pay for input and output tokens.
- Russian is "pricier" than English. The same meaning takes more tokens in Russian — you burn through limits faster.
The context window
The context window is the maximum amount of text the model holds "in its head" at once: your messages, its answers, uploaded files, system instructions. Modern Claude models have a window of hundreds of thousands of tokens — hundreds of pages of text.
The window looks enormous, but it fills up faster than you expect: a long chat with big documents and detailed answers piles on tokens with every reply, because the model rereads the entire history at each step.
What happens when it overflows depends on the interface: the chat may refuse new messages, "forget" the start of the conversation, or compress old replies into a short summary. Either way, details from the start of a long dialogue gradually lose their influence on the answers.
A symptom of an overflowing context: the model starts contradicting agreements from the start of the chat, forgets rules you set, mixes up details. The cure is a new chat with a short recap of what matters.
The knowledge cutoff
The model's training ends at a specific moment — the knowledge cutoff. Anything that happened after that date the model does not know: new events, exchange rates, software versions, yesterday's news.
Ask about an event past the cutoff and you get one of:
- an honest "my information ends on such-and-such date";
- an answer based on outdated data, with no warning;
- a plausible fabrication.
Many interfaces (Claude included) can search the internet — then the model pulls in fresh data and answers from it, with links to sources. The rule is simple: for any time-sensitive facts, either turn on search, or give the model fresh data yourself, or verify the answer by hand.
Practical consequences
- A big document beats a "from memory" question. Instead of "what does law X say," attach the text of the law — the model answers from the document, not from recollections of its training.
- One chat, one task. Long catch-all chats overflow the context and degrade answers.
- Put what matters near the end. In a very long dialogue, restate the key instructions in a fresh message.
- Check dates and numbers. Anything that could have changed after the cutoff is suspect by default.
Key takeaways
- The model counts text in tokens; limits and prices are measured in them, and Russian text is "pricier" than English.
- The context window is everything the model sees at once; long chats gradually lose details from the start.
- The model's knowledge ends at the training date; fresh facts come via web search or by hand.
- An attached document is more reliable than the model's memory.