grok 4.7, Jev and Opus 5.5 split roles to cut cost per task by 84%
The author uses Jev only to make choices (picking notes, routing tasks, choosing recovery paths and checks), Opus for hard thinking and grok for the work, cutting cost per finished task by 84%.
grok 4.7 + jev + opus 5.5 = the smartest thing i have run yet i call it nitro. three models, one job each, and only one of them ever writes a word it cut my cost per finished task by 84% and the time by a bit more than half jev is the whole trick. it does not write anything. it only decides: → which project notes go into the context and which stay out → which tasks go to the cheap worker instead of opus → what to do when a tool call fails → which checks to run before the full test suite opus 5.5 does the hard thinking, but only on what jev let through. grok 4.7 does the work: the browser, the files, the terminal jev costs $0.042 per million tokens in and nothing at all on output. on a full run it was under 1% of the bill it answers 25 questions at the same time and comes back in about a quarter of a second i had it wrong for two days. i let opus pick its own notes and it picked all of them, every time turns out the big model does not know how to leave things out. that is the job you hand to the cheap one jev cannot make an answer up. it only picks from the list you allow it those four questions are the whole build. More about that in article below ↓
This post quotes another post: x.com/razeden0/status/2101981166500815075