我不断撞上 AI 编码工具的“配额墙”,于是我构建了一个路由器
📄 中文摘要
使用 OpenClaw、Codex、Cursor 或 Claude Code 的用户常常会发现,自己的高级模型配额在一周中途就用完了。通过分析自己的使用日志,发现大部分提示并不复杂,而是一些乏味的实用工作,如“总结段落”、“重新格式化 JSON”、“解释错误”和“转换为 TypeScript”。这些简单的请求并不需要高端模型,但却被发送到这些模型,导致配额迅速消耗。经过一周的日志分析,发现使用情况的分布相对稳定:约 55% 的请求属于简单类型,25% 为中等复杂度,15% 为复杂请求。
📄 English Summary
I kept hitting the "quota wall" with AI coding tools. So I built a router.
Users of AI coding tools like OpenClaw, Codex, Cursor, or Claude Code often encounter the issue of their premium model quota depleting midway through the week. An analysis of usage logs revealed that most prompts were not complex but rather mundane utility tasks such as 'summarize this paragraph', 'reformat this JSON', 'explain this error', and 'convert this to TypeScript'. These simple requests do not require high-end models but were being sent to them, consuming quota rapidly. After a week of logging, the distribution of usage was consistent: approximately 55% of requests were simple, 25% medium complexity, and 15% complex.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等