TL;DR
The short version
TypeSafe's Jev returns Choice, Score, or Noul at about $0.04 per million input tokens — built for decisions, not prose.
Claire Vo clusters PRs for cents, triages mail and comments, and unlocked a ChatPRD insights graph by putting Jev in front of frontier models. Pattern: Jev alone is okay; Jev with an LLM buddy is powerful.
Based on How I AI I'm using Jev more than Opus 5.5 or GPT-6. Here's why. (28 Sep 2026).
Typed answers, dirt-cheap
While Opus 5.5 and GPT-6 launched, Vo opened How I AI with Jev: text in, type-safe values out. Choice from a list, Score on a named scale, Noul as probability a statement is true. About $0.04/MTok in; output free. She spent under $10 in a heavy week.
Jev is really great for making decisions. […] I've been calling this […] a decision model.
Claire Vo
Use it for smart if/else, classification, and real-time loops. Leave generation to generative models.
Piles: PRs, sessions, mail
Pairwise "are these related?" over GitHub PRs, then a cheap generative model to name clusters: ~112 marketing PRs for about a cent; ~1,700 ChatPRD PRs / ~17k pairs for about nine cents — enough for board-shaped effort splits she could sanity-check.
Same pattern on local Claude Code/Codex session logs and Gmail subject+snippet delete scores. Filter and tag first; apply expensive agents only to the interesting buckets.
Jev in front, brain behind
Jev alone is okay. Jev with an LLM buddy is super powerful.
Claire Vo
A ChatPRD product-insights graph that burned thousands on frontier-only prototypes became affordable when ~1,100 signals ran through on the order of 200k Jev judgments (~$4) before Astra-class analysis. Afternoon demos: 4,500 podcast comments into sentiment + episode-idea boards, and a live voice→emotion-color→quote toy.
Copy the habit: write answers in advance, run them over every item, hand only the interesting buckets to a generative model, keep a human for anything board-facing.
