A long chat re-sends its history every turn, and caching only covers the unchanged prefix
Long chat sessions incur costs not from the current message, but from re-sending the entire conversation history with each turn. For instance, a two-line reply on turn 40 includes the previous 39 turns, making the bill reflect the full thread despite the on-screen appearance. A larger context window doesn't solve this; it merely encourages longer threads, increasing costs. The suggested solution is to start new sessions when tasks change, retaining old ones only for the same task, balancing cost with workflow friction.
This account uniquely details how a long chat session's cost is tied to re-sending full history, unlike the common misconception that only new messages incur charges.
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年10月10日 00:44 UTC
收录当时偏移：UTC+02026年10月10日 04:00 UTC
- 发布
- 2026年10月10日 00:44
- 收录
- 2026年10月10日 04:00
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
讨论趋势
百分比基于采集到的讨论信号，不代表新增评论数或独立参与人数。曲线仅用于同一话题在不同时段的比较。
The cost of a long session isn't in the message you just typed. It's in everything before it going back through the model again. A two-line reply on turn 40 carries the previous 39 turns with it. On screen the session looks light. The bill reflects the whole thread.
Prompt caching softens this, but only while the prefix stays byte-identical. Edit an earlier message, or let the framing drift even slightly, and everything after the edit point is a cache miss. You pay full price for history you already paid for once. This is the part that produces the mismatch people describe: a session that felt cheap on screen and came back with a number that doesn't match the work they think they did.
A bigger context window doesn't fix this. It makes staying in the same thread more comfortable, which extends the thread, which is the direction the bill doesn't want. The fix is unglamorous: start a new session when the task changes, and keep the old one only while it's still the same task. Whether that's worth the friction depends on how often your threads actually change direction mid-session, and I'm not sure that ratio is the same for everyone.