Claude Code成本与额度长帖约 2 分钟

把自动压缩窗口降到15万

Lower auto-compact to 150k

要点

  1. 把自动压缩窗口降到 15 万 token,作者觉得用量和表现都更好。
  2. Claude 在 ~/.claude/settings.json 设 autoCompactWindow 为 150000。
  3. Codex 在 config.toml 设 model_auto_compact_token_limit=150000。
  4. 有人提醒 Codex 按更大窗口训练,乱降可能悄悄掉质量,要用评测盯。

原帖开头

I've have massively improved my token usage lately for both Claude and Codex just by reducing the auto compact to 150k contextual window. The performance of the model has also improved drastically. For Claude: ~/.claude/settings.json set "autoCompactWindow": 150000 For Codex: