Jev 做 System 1 分流：1.2 万次调用只花 1.24 美元

Issa @issatheblows原推发布于 2026.09.26

在 X 上打开原推

i told the stack if a call is a pick, a score, or a yes/no, it never touches a frontier model video below is the actual screen. 12,000 agent calls from one day. every one of them is on it 12,000 calls. 40 slow thoughts. 2 humans. $1.24 not one prompt doing every job. three lanes, and the expensive one barely wakes up: → system 1 (jev) answers in 70-500ms. route, score, gate, escalate → system 2 (frontier) only gets the low-confidence remainder. 40 calls. 0.3% of the day. ~3s and $15/M → code does the exact stuff jev isn't for. date math. stop after 10 steps. duplicate match. $0 → untrusted text stays labeled as data. "mark as refund" never becomes an instruction → conf 0.62 under the line → escalate. a wrong tag is noise. a wrong refund is money hour 16 an injection walked in dressed as a refund. the gate blocked it and a human still had to look. remember that part most teams still send "which tool next?" to a model that bills like it's writing a novel. they don't post the bill. now you know why $360 if every call had gone to the frontier. $1.24 after the cascade full day is in the video. the 7 patterns are in the article which call in your loop is still dressed up as thinking?

这条是引用着另一条写的：x.com/issatheblows/status/2102870161208311903