用 Jev 做路由层搭 24/7 常驻 agent 阶梯

starmex @starmexxx原推发布于 2026.10.03

在 X 上打开原推

sam altman's openai just made gpt-6 astra run 24/7 inside openai dots, and the cheapest way to keep it running is a $0.042 model from typesafe called jev that can't write a single word leo varga built the whole stack in one weekend: openai agents sdk builds the agent in 11 lines -> jev routes every decision -> gpt-6.1 sol drafts the routine work -> gpt-6 astra only wakes up for the hard cases -> leo approves what can't be undone jev is the strangest model release of the year: it can't chat, can't code, can't write a report. it answers yes or no, pick one, or rate it, with a confidence score on every answer, and it reads at $0.042 per million with zero output cost. a million decisions cost about $42. harrison chase's langchain already ships a router where jev picks the cheapest model that fits the task what happens to leo's 400 overnight emails: jev reads all 400 and sorts them for about two cents -> 380 get ignored, 12 go to gpt-6.1 sol for drafts, 8 get flagged for leo -> a 90-page client contract rated complex escalates straight to gpt-6 astra -> every risky action like delete_customer passes jev, then hard-coded rules, then leo's approval queue -> at 07:00 leo reads 8 flags and 12 drafts in 15 minutes instead of 2 hours sam altman prices gpt-6 astra at $10 per million read and $50 per million written. gpt-6.1 sol costs $2 and $10. jev costs $0.042 and nothing for output. most people will plug the biggest model into an always-on agent and switch it off after one bill. leo built the ladder first, and now he wakes up to finished work every morning

这条是引用着另一条写的：x.com/noisyb0y1/status/2105964235087856080