What's this about?
Three practical findings from a heavy user in the first 24 hours. Disable the 1M context window — set `CLAUDE_CODE_DISABLE_1M_CONTEXT: "1"` in `~/.claude/settings.json`; the effect is especially noticeable on Fable 5.1. Avoid Max effort — it costs 1.5× or more, which is why he usually picks Fable high or Opus high, and the mobile and desktop clients now warn you when you select it. Be sceptical about `/compact` — Anthropic advises that with cache reads much cheaper you need not compact frequently, but his own testing found this worse than 200k auto-compaction, and if the cache has already expired (one hour on Claude Code subscriptions), compacting becomes expensive.
Why does a larger context window cost you more?
Because context is re-sent every turn, and a larger window means the model keeps more rather than compacting. You pay for those tokens on every subsequent step, so the window that lets you avoid compaction is also the window that makes each step more expensive. Whether that is worth it depends on whether the retained context is being used — and in most sessions, most of it is not.
Why does this matter to you?
- These are the first-day empirical findings that do not appear in any announcement.
- The `/compact` finding directly contradicts vendor advice, tested rather than asserted, with the caveat about cache expiry that makes both true in different conditions.
- The client now warning you about Max effort is a small admission that the cost difference is significant.