Home / What a language model is, without the jargon
The token and the window, what the model actually sees
The model sees tokens stacked inside a window with a fixed ceiling, not whole words taken one by one, and a conversation approaching that ceiling changes content before it even reaches its maximum cost, since Claude then summarises the oldest exchanges.
Claude does not read words: it reads tokens, small chunks of text, sometimes a whole word, sometimes part of a word or a punctuation mark. Everything it can see at a given moment is measured in tokens, the question asked, the previous exchanges in the same conversation, and the documents pasted or attached. This whole set fits inside a window whose size is fixed, not inside a memory that would expand as the exchange goes on.
What the window actually contains
On paid plans, this window reaches one million tokens for the Claude Fable 5, Claude Opus 5 and Claude Sonnet 5 models, and two hundred thousand tokens outside these three models, according to the official help centre consulted on 2 September 2026, a figure that this lesson's figure cites along with its source. A share of this window is always reserved for Claude's answer, according to the same help centre, so the longest possible conversation is a little shorter than the whole window. The conversation interface does not display this figure while you are writing.
What happens when the window nears its ceiling
The practical consequence is stated in a rule published by Claude's help centre: when a conversation nears the limit of its context window, Claude summarises the oldest messages to make room, and this summary does not count against the account's usage limit. The conversation therefore changes content before it even reaches its ceiling: the very first exchanges stop being read word for word, they get replaced by a condensed version. An instruction given very early in a long conversation can thus end up summarised rather than read as written.
Estimating without a displayed counter
In the absence of a counter built into the conversation screen, a rule of thumb published by Anthropic makes it possible to estimate a token count from a text's number of characters. This lesson's figure gives this figure along with the two window sizes.
Pasted text: about 800 characters, measured with the word processor's counting tool
Estimate: 800 divided by the figure given in the diagram, so about 200 tokensKeep this estimate in mind before pasting a long document at the start of a conversation. That is the exercise this lesson proposes, and it sets up the lesson on the order of information in a long prompt, where the position of a sentence matters as much as its length.
The stack that fills up the context window
Two window sizes, and the rule for estimating a text
An association pastes a twelve page report into the same conversation, then the exchanges from the three previous weeks, before asking its question. Claude answers by first summarising the oldest messages in the conversation before dealing with the question asked.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: The conversation was nearing the limit of its context window at the time of this answer, since this automatic summarising of old messages is triggered when the ceiling is approached.
What this does not establish: That the twelve page report was cut off or lost, which rereading the pasted document would confirm.
The three most common miscalibrations
- Too broad Claude has reached its absolute limit and can no longer answer any question until the conversation is deleted.
- Too narrow This summary says nothing about the state of the context window, it is simply a stylistic choice for the answer.
- Beside the point This situation shows that pasting a twelve page report at the start of a message is a good practice worth repeating systematically.
- Claude does not read whole words but tokens, fragments that can split a word into several pieces.
- The size of the context window depends on the model chosen, it is not the same for every Claude model available on a paid plan.
- When a conversation nears the ceiling of its window, Claude summarises the oldest messages, which then stop being read word for word.
- Estimating a text's token count is done from its number of characters, a rule of thumb used in the absence of a direct display while writing the message.
Take a long text you have already sent to Claude, count or estimate its number of characters with your word processor's counting tool, and divide this total by the figure given in this lesson's diagram to estimate how many tokens it takes up.
These points depend on an interface or a rule that may have changed since this was written. Check them on your own screen before relying on them.
- Check on your own interface whether a token counter or a fill gauge now appears near the input field, this display can change depending on the plan and product updates.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic, context window size by model and plan consultée le 2026-09-02
- How do usage and length limits work? consultée le 2026-09-02
- Pricing (rule of thumb on tokens) consultée le 2026-09-02