The Context Window as a Budget

Suppose Forge, a coding agent that works on a repository, sends 6,000 tokens to the model on its first step. On step 20 it sends 40,200 tokens, and on step 40 it sends 76,200. The work of step 40 is no larger than that of step 1. Over all 40 steps, Forge sends 1,644,000 tokens, and only 76,200 of them are different tokens.

Those numbers describe one resource with three separate limits: how much text fits in a call, how much of it the model reads accurately, and the cost of the repeated text. Where does a reduction save the most?