WHAT NEW IN X11.06 FRI

AI collections from X · 23 collections · 69,408 posts

jevVIDEO

Routing agent calls through Jev cuts a day of 12,000 calls to $1.24

Picks, scores and yes/no calls go to Jev, and only the 0.3% low-confidence remainder escalates to a frontier model, bringing a day's cost from $360 to $1.24.

Issa
@issatheblowsArchived 2026.09.26Originally posted 2026.09.26
Issa @issatheblowsOriginally posted 2026.09.26
Open original on X

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?

This post quotes another post: x.com/issatheblows/status/2102870161208311903

x.com/issatheblows/status/2103582665077035231Archived 2026.09.26 10:13