confirmed: you probably shouldn’t use claude code’s full 1m context window.
two reasons:
- compacting a massive context is expensive and burns through your token quota.
- output quality gradually degrades as context fills up. the model starts missing details and losing track of earlier instructions.
400k tokens seems to be the sweet spot between cost and output quality.
you can configure claude code to automatically compact at 400k by adding this to ~/.claude/settings.json:
{ "autoCompactWindow": 400000 }
400k isn’t a magic number, but there’s little point in filling up 1m tokens just because you can 😁