HiRoute brings staged model routing to local AI coding
Developer Cheng Tan describes HiRoute, an open-source AI gateway that runs on a developer’s machine and selects models by task stage while preserving a continuous coding session. In a controlled Higress concurrency-defect test, the reported routing setup passed all three runs, versus one of three for Qwen 3.8 Flash alone, while the author says its estimated API-equivalent cost was 59% below using GPT-6 Astra throughout. The project, testing, usage, and cost figures are first-person claims from a single report.
HiRoute offers a concrete local workflow for reserving stronger models for difficult decisions while assigning routine testing and follow-up work to lower-cost models under shared acceptance checks.
作者:澄潭
整理本机环境时,我打开 LoongSuite Pilot 看了眼近 30 天的用量。面板上是 30020.4M token,约合 300 亿。
LoongSuite Pilot 的本机用量总览。近 30 天共 500 个会话、150,578 次模型请求和 149,656 次工具调用。
如果还停留在一问一答写代码片段的阶段,300 亿很难想象。把一个缺陷交给 Coding Agent 后,它要读代码、查调用链、改实现、跑编译和测试,遇到报错再改一轮。
我的记录里,一个会话平均要发出约 300 次模型请求。这 300 亿 token 中,约 98% 是长上下文的缓存读取。
#HiRoute#Higress#Alibaba Cloud AI Gateway#GPT-6 Astra#Qwen 3.8 Flash#Codex#Claude Code#LoongSuite Pilot#Pi#DeepSeek Harness