WHAT NEW IN X11.01 SUN

AI collections from X · 22 collections · 63,170 posts

jevVIDEO

The cost case for replacing LLM calls with Jev for yes/no decisions

A comparison finds 10,000 binary decisions cost 42 cents on Jev versus $40 in input alone on Opus 5.5, plus three questions and rollout steps for moving decisions to Jev.

Ryven
@imryvenArchived 18 days agoOriginally posted 2026.10.12
Ryven @imryvenOriginally posted 2026.10.12
Open original on X

I still don’t understand why people keep paying frontier-model prices to answer yes or no 10,000 yes/no decisions, about 1,000 tokens each → Jev, a model that can’t write a sentence: 42 cents → Opus 5.5: $40, and that’s input only that’s about 95x cheaper. an agent making 100,000 of these a day pays $126 a month instead of $12,000 most of your agent’s calls were never writing. they’re a surgeon checking temperatures, because the surgeon was the only one on staff $0.042 per million input tokens, output free. up to 90% off real agent bills and it doesn’t answer with a sentence. it answers with odds → ship 52% → return 46% → confidence 0.18 ship won by six points at 0.18 confidence. an LLM would write “looks good to ship” and hide how close the call was three questions tell you which of your calls to move off the LLM: → can you list every valid answer right now → would a person glance at it and just know → does it happen often enough to matter three yeses, move it. one no, leave it how to start without breaking anything: → pick one low-risk decision where every answer has a name → write what belongs in each option before you touch the model → run it in shadow mode next to your current logic for a couple of days. change nothing → set thresholds from accuracy against confidence, not from a blog post → automate the safest branch first, keep a human on the uncertain ones one rule that saves you: hard rules in code outrank the model. an irreversible change escalates no matter how confident the verdict is it can’t invent an option, but it can confidently pick the wrong one. and never ask a model whether tests passed. read the exit code go count the pauses in your last agent run. not where it wrote, where it chose. that’s the surgeon, still checking temperatures reply with the one decision in your agent you’d move first full setup, code and rollout in the article below

This post quotes another post: x.com/imryven/status/2104255840819491030

x.com/imryven/status/2109631075978707309Archived 2026.10.13 10:08