What matters is context, and how saturated the session gets. The histogram bar tells you when it's worth clearing and starting fresh: efficiency is high early on, then it degrades and you burn more tokens for the same work.
I tried to cut my Claude Code bill by asking for shorter answers.
Then I measured five of my biggest sessions.
Output was ~10% of spend. Prose was a quarter of that.
I'd been optimizing 1.5% of the bill.
One trap: "percent of context window" is the wrong meter.
500K in a 1M window reads as half full.
It costs 2.5x what 200K costs β every turn, until you clear.
Absolute tokens are what maps to money.
@TgLa7 Sara' interessante vedere quanti voti prendono i candidati renziani di camera e senato x capire la presa che ancora ha l'ex segretario sui deputati e senatori.