Skip to content
AI Training

What the model can see

You have been working with a chat assistant on a report for two hours. In your first message you said the report is for the finance team and must stay under two pages. Now you ask for the closing section, and the answer runs to four pages and addresses “the engineering team”. The transcript below is illustrative, written for this lesson rather than recorded from a model, but the pattern is one you will meet.

Prompt (illustrative, not a recorded transcript)

Write the closing section now. Same audience and length as we agreed.

Response

Here is the closing section for the engineering team. I have kept the technical detail you asked for, so it runs to about four pages. Let me know if you want the deployment appendix as well.

The model did not forget in the way a person forgets. It never saw your first message when it wrote this answer. In the previous lesson you learned that a model reads tokens and predicts the next one from everything it has seen so far. This lesson is about the limit on “everything it has seen so far”, the context window. In this lesson we measure a page of text in tokens and work out how many pages a window holds, then watch a chat grow past it.

Count the tokens of a page

Section titled “Count the tokens of a page”

A context window is measured in tokens, so the first step is to get a feel for how many tokens a page of text is. The counter below contains a sample of meeting notes, about half a page. Press Count.

      The first part of the report should read:

      characters: 460
      words: 80
      tokens (estimate): 115

      The counter says “estimate” for a reason. It applies the common rule of thumb for English, about four characters per token. A word of up to six letters counts as one token and a longer word as one token per four letters, with each punctuation mark counted as a token of its own. Real tokenizers are more detailed than that, and every vendor has its own, so the exact count for a model comes from that vendor. A call to the model through the vendor’s programming interface (API) returns the number of tokens it used along with the answer, and vendors publish a token-counting call or a web page that runs their tokenizer. For planning, the rule of thumb is close enough: a full page of English text is a few hundred tokens.

      How a tokenizer splits text

      Section titled “How a tokenizer splits text”

      The counter cuts “delivery” into two tokens by its four-letter rule. That rule is a stand-in for what a real tokenizer does. A real tokenizer has a fixed vocabulary of tens of thousands of chunks learned from training text, and it cuts where that vocabulary happens to have a chunk. A common word like “delivery” is in the vocabulary whole and costs one token, however long it is. A rare word or a name is split into pieces instead, often at places that have nothing to do with syllables or spelling. “Chidi” from the sample is two tokens in current tokenizers, and a word in another language may end up as one token per letter [1]. The counter cannot know which words are common, so it applies one rule to all of them.

      Asking a model how many letters “strawberry” has goes wrong so often for the same reason. The model receives the token for “strawberry” as one number, and the letters are not in its input, so it predicts a count from what it has seen written about the word, the same way it predicts any other answer. Counting characters yourself, as the counter does, is the reliable way.

      How many pages fit

      Section titled “How many pages fit”

      The second part of the counter’s report takes the token count from the first part as the size of one page, and shows how many such pages fit in windows of different sizes. For the sample it reads:

      one page: 115 tokens
      8,000-token window: 69 pages
      32,000-token window: 278 pages
      200,000-token window: 1,739 pages
      1,000,000-token window: 8,695 pages

      The window sizes are round numbers that stand for the sizes on the market in 2026. The 8,000 row is the size of a few years ago, kept for the contrast. Small models that run on a laptop have windows at the low end today, and the large vendor models offer 200,000 tokens or more [1]. At the top end the window holds a few thousand pages of text, which sounds like more than you will ever need. The next section shows why it still runs out.

      A big window is not free to fill. The model processes every token in the window for every response, which makes a fuller window cost more per answer and take longer to answer. Nor does the model weigh every part of a full window equally well: a fact in the middle of a long paste counts for less in the prediction than the same fact at the top or at the bottom, as the previous lesson noted under instruction dilution [2].

      What the window holds

      Section titled “What the window holds”

      The window is not only your messages. For one response, the model sees all of this at once, and the total must fit:

      • The instructions the product or your team set up, which you may never see. In a chat product these are often a few pages by themselves.
      • The conversation so far, with every message from you and every answer.
      • Everything you pasted or attached, in full. A document you attach is turned into tokens like any other text.
      • Results from tools the assistant used, such as a web search or a file that it read.
      • The answer it is producing, token by token, since each new token is predicted from the whole window including the answer so far.

      From this follows the one rule of the lesson: anything outside the window does not exist for the model. Yesterday’s chat, the file you did not paste, and the company wiki are all invisible unless something puts them into the window for this response [2]. The same rule explains why a model “remembers” nothing between conversations, because the weights do not change when you type and so the window is the only place a fact can be.

      Checkpoint · sort

      What can the model see?

      Section titled “What can the model see?”

      When the early facts fall out

      Section titled “When the early facts fall out”

      Your chat with the assistant grows with every turn, and so does the text sent to the model for each response. A long chat with pasted documents can pass a 200,000-token window in an afternoon. Products handle the overflow in different ways. Some silently drop the oldest turns, some replace them with a short summary, and some refuse to continue and ask you to start a new chat. In the first two cases you get the report chat from the top of this lesson: the answers are still fluent, but a fact from the start of the chat has left the window.

      Restating the facts is the fix in the moment. To prevent the problem, keep the facts the model needs every time in one message you can paste again, and start a fresh chat for a new piece of work instead of adding it to a long one.

      Checkpoint · choice

      Why did the answer change?

      Section titled “Why did the answer change?”

      After two hours of chat, the assistant that had been writing for the finance team in two pages suddenly writes four pages for the engineering team. Which explanation fits what you know about how a model produces an answer?

      Exercise

      Copy about one page of your own writing, an email or a report section. Replace the meeting notes in the counter above with it, and press Count. The counter runs in your browser and sends your text nowhere. Write down the three numbers: characters, estimated tokens, and how many of your pages fit in a 200,000-token window. With these numbers you can judge whether a document fits before you paste it, and after how many turns a chat is full.

      A good result is one line like “my page is 3,100 characters, about 750 tokens, and 266 pages fit in 200,000 tokens.” Then answer this: how many pages of instructions, pasted documents and earlier turns would your usual chat sessions need before the first message falls out of the window?

      Stretch: Paste the same page into a chat assistant and ask it how many tokens the page is. Compare its answer with the counter, and decide which one you trust and why.

      Recap

      1. A context window is the maximum number of tokens the model takes into account for one response, and a page of English text is a few hundred tokens [1].
      2. The window holds the instructions, the conversation so far, anything pasted or attached, tool results and the answer being produced. Anything outside it does not exist for the model [2].
      3. When a long chat contradicts itself, an early fact has left the window. Restate the fact in your next message instead of scolding the model.
      4. A fuller window costs more per answer, because every token in it is processed for every response, and the model does not use every part of a full window equally well [2].

      You can now

      • Explains tokens, context and sampling in plain words

      References

      Section titled “References”
      1. Addy Osmani, Ivar Soares Urdalen, Leo Simons. How agents think: tokens and context, reasoning strategies, model choice, system prompts. Agent Engineer Course. Course. AEC-02
      2. Anthropic. AI capabilities and limitations. Claude Academy. Course. Academy ai-capabilities-and-limitations