The improved token efficiency would only account for the output and reasoning tokens occupying context, right? What about the user messages and other context like file contents? I’m not sure how these measure up in a long conversation
replyFor agentic use it impacts input tokens because the output is fed back into the model as input on subsequent turns.
reply