reddit · SideProject · 10-08 17:50 69 分 ← 返回列表

I found 56% of my Claude usage was re-reading my own context, so I built a tool to catch it

I built optimAIzr, an open-source AI usage optimizer for Claude Code + Codex. I work as a senior SWE so I got curious where my Claude usage was actually going. In my own logs, 56% of the cost was Claude re-reading my existing conversation (which I shared here last night as well). Coming back to long sessions after the cache expired was especially painful: I did it 35 times in 27 days, and those first messages cost $98.74 at API prices vs $2.95 with a warm cache. So I built

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：Claude Code / Codex 用户在长会话中因缓存过期被迫重读上下文，产生大量隐性 token 成本却难以察觉和量化。

目标用户：高频使用 Claude Code / Codex 的个人开发者与 AI 编码团队（尤其按 API 付费或关心用量上限的人）。

证据：56% of the cost was Claude re-reading my existing conversation... I did it 35 times in 27 days, and those first messages cost $98.74 at API prices vs $2.95 with a warm cache. / I also cut my own claimed savings from $121/mo to $13/mo after checking the numbers again.

变现：开源 CLI 免费获客 + 付费团队版 Dashboard（多账号聚合、成本归因、月度节省报告订阅），或按席位订阅的 CI/团队用量治理服务。

MVP：2 周内做一个本地 CLI：解析 Claude Code/Codex 日志中的 token 与 usage 元数据，输出会话成本排行、缓存即将过期实时提醒、以及「避免浪费」的节省金额估算，无需账号、纯本地。

竞品：ccusage、Claude Code Usage Monitor、CCSeva

摘要：一位资深 SWE 发现自己 56% 的 Claude 成本花在重读自身上下文上，27 天内因缓存过期重开长会话 35 次，首次消息按 API 价需 $98.74，而热缓存仅 $2.95。他据此做了开源工具 optimAIzr，本地读取用量元数据并提供缓存过期预警、会话成本排行与逐条成本提示。数据具体、痛点真实且已被自身验证，但原帖为免费开源发布，未出现他人付费意愿原话。