A coding agent keeps your whole conversation in its head. When that gets too big, it writes a summary and carries on from the summary. Compact too early and you keep paying to rebuild the same context. Compact too late and every request drags a heavy load. These charts show where the sweet spot sits, using real sessions.编程助手会把整段对话都放在脑子里。对话太长时,它会写一份摘要,再从摘要接着干。压缩太早,就要一遍遍花钱重建同样的上下文;压缩太晚,每次请求都背着沉重的负担。这些图用真实的使用记录告诉你,平衡点大约在哪里。
Each page has a chart and a few ready-made picks for the trigger setting.每个页面都有一张图,还有几个现成的触发点建议。
Codex sessions on DeepSeek Flash.Codex 里 DeepSeek Flash 的会话。
Cheapest near a 124K trigger触发点在 124K 附近最省钱
Open the chart查看图表 → Claude CodeClaude CodeClaude Code sessions on Opus 5.5, main conversations and subagents.Claude Code 里 Opus 5.5 的会话,含主对话和子代理。
Cheapest near a 120K triggerautoCompactWindow: 153000触发点在 120K 附近最省钱autoCompactWindow: 153000
Claude Code sessions on Sonnet 5.5, timing borrowed from Opus.Claude Code 里 Sonnet 5.5 的会话,耗时借用自 Opus。
Cheapest near a 104K triggerautoCompactWindow: 137000触发点在 104K 附近最省钱autoCompactWindow: 137000