Jev, a decisions-only model, matches gpt-4o-mini 50/50 while being 2.4x faster and 3.4x cheaper

Al_Grigor · x · 2026-09-18

A new model called Jev doesn't generate text — it only makes decisions. You send a state (text or JSON) plus typed questions, and it returns structured JSON with probability distributions.

The author replaced gpt-4o-mini for routing tasks across his 3 agents and benchmarked both on 50 hand-labeled requests: identical 50/50 accuracy, but Jev was 2.4x faster (0.32s vs 0.76s median) and 3.4x cheaper.

It's aimed at anywhere you pay a cheap model for structured-output decisions: routing, guardrails, triage. Full write-up with API and benchmark included; author is running a workshop Tuesday 16:30 Berlin time for AI Shipping Labs members.

Related event: TypeSafe Launches Jev, a Decision-Only Model That Never Writes a Word(44 posts)→

Original post →

More from coding & agent

coding & agent channel →