Anthropic engineer routes models through Jev, cutting the bill 85%
Claude splits tasks and Jev scores and picks the model; Opus handles only 19% of hard tasks but takes 60% of the cost, dropping a week from $1,940 to $287.
ANTHROPIC ENGINEER LEAKED HIS ROUTER CONFIG WHERE A 4-CENT MODEL BENCHES OPUS 5.5 ON 81% OF TASKS AND THE BILL DROPS 85% $1,940 used to buy one week of work, and now $287 buys that week with more output the file has one rule: he never picks the model Claude splits the job → jev scores each task → one model does it → back to Claude jev sits in front of all three - it reads scope, doubt and what failure costs, and never writes a word Opus 5.5 runs on high for 914 of 4,812 tasks - the hard calls - and takes 60% of the bill GPT-6 Astra runs on medium for 3,106 tasks - drafts and code - and takes 29% Grok runs on medium for 792 tasks - anything x said an hour ago - and takes 11% Opus did 19% of the work and ate 60% of the bill - the 4-cent model decides where $171 lands Opus wrote essays for 1-line jobs - with jev in front, a task fell from 40 cents to 6 4,812 tasks → 914 hard calls → 37 stops for your yes → 0 moves you did not approve it does not make Opus smarter - it stops you paying Opus prices for grunt work remove jev and tomorrow Opus writes an essay for every one-line task again Also you can merge it with Grok and receive Nitro ↓
This post quotes another post: x.com/razeden0/status/2101981166500815075