Home / Never running out of tokens

Understanding the token and the context window

Every word read or written consumes a token, and reopening an old session reloads, and therefore repays, its whole context window.

Claude does not read whole words but tokens, small fragments of text. Every question asked, every answer received and every file attached is first split into tokens before being processed, and it is this splitting that determines what is deducted from your quota and what fills the conversation's working memory. On a subscription plan, the one most organisations follow, the token is not billed per unit: it consumes a quota, which does not make the question any less concrete, an exhausted quota stops the work just as surely as a refused payment. Some organisation plans are instead billed by consumption, and there the count does become an actual expense: whoever manages your accounts knows which one you have.

The context window, a limited working memory

The context window is the space that holds everything Claude can see at once to answer: the message you have just written, the full history of the current conversation, and the files you have attached. This space has a ceiling, and that ceiling varies with the model used, the figure below gives the measured values. Once the window is full, there is no more room to continue without making a choice: summarise, clean up, or start again on a lighter base.

Reopening an old session reloads everything

The most common trap for someone discovering the tool fits in one sentence: picking up the thread of an already long session does not only cost the new message being written, it reloads everything that has already been exchanged before it. A session open for several days, loaded with dozens of exchanges and several files, therefore recounts its entire history with every new question, even a short one. This is the mechanism that explains why an old session gradually becomes slower and heavier on quota, with nothing appearing to have changed on the user's side.

What actually piles up

A simplified example shows what accumulates in a session left open for several days:

Jour 1 : questions 1 à 20, plus un fichier joint
Jour 2 : la session reprend, questions 1 à 20 rechargées en entier, puis questions 21 à 45
Jour 3 : la session reprend encore, questions 1 à 45 rechargées en entier, puis questions 46 à 60

Every resumption puts everything that came before back in front of Claude, never just the last question asked. One nuance matters for your quota: Anthropic's documentation states that content attached to a project is cached, and that only new or non cached portions are counted. The weight remains real, but it is not paid twice over identically. It is this accumulation that makes an old session heavier to carry than its content for the day would suggest.

The rest of this module explains what to do with this observation: what to do once the limit is reached, see what to do when the limit is reached, and how to spread out your work to avoid hitting it too often, see the right habits for saving.