WHAT NEW IN X10.28 WED

X 上的 AI 合集档案 · 22 个合集 · 58,924 条

jevVIDEO

用 Jev 做模型路由,月省 36% LLM 账单

有人把每月 31 万次请求交给 Jev 分档,按难度分流到 Haiku/Sonnet/Opus,路由成本仅 4.68 美元,账单降 36%。

Blanc
@scrblanc16 天前收录原推发布于 2026.10.11
Blanc @scrblanc原推发布于 2026.10.11
在 X 上打开原推

TYPESAFE'S JEV ROUTED 310,000 REQUESTS A MONTH TO THE RIGHT MODEL AND CUT THE LLM BILL 36% FOR $4.68 OF ROUTING the router costs $4.68 a month, and the bill it cuts is $1,535.40 request → jev → tier → answer → log the router on Jev 1.13 sorts 310,000 requests into easy, medium and hard for $4.68, the easy 241,800 go to Haiku 4.5 for $1,668.42 the medium 58,900 go to Sonnet 5.5 for $812.82, the hard 9,300 go to Opus 5.5 for $256.68 none of them runs on “latest”, because a router that changes its mind mid-month changes the bill and the answers together 310,000 routing decisions a month, 10,333 a day, and a person triaging prompts by hand does about 400 a day that is a team of 26 doing nothing but deciding which model gets which question one request costs $0.0138 on Sonnet 5.5 alone and $0.0088 routed - that is 1.6x cheaper, the smallest multiplier in this series and the one that pays every month and here is the honest part of this file: if routing accuracy is anywhere near the 68% on TypeSafe's own chart, about 99,200 of the 310,000 land on the wrong tier a wrong tier down costs you quality and a wrong tier up costs you money, so anything under 0.80 confidence goes one tier up $1,535.40 a month is 5.1 customers at $299, $18,424.80 a year routing saves 36%, and the 3% of requests that reach Opus are the ones your customers remember save this and paste it into Claude Code - let it route your 310,000 requests for you ↓

x.com/scrblanc/status/2109177369466450133采集于 2026.10.12 10:05