JEV, a no-text decision model claims 200x speed and 400x cost savings, with an early port of Claude Code compaction hitting 4,575 stars in five days
gaganghotra_ · x · 2026-09-21
Key facts
TypeSafe's Jev is a decision-only model that generates no text: you hand it state and typed questions, it returns a choice from a set you define plus a calibrated confidence score — it physically cannot return something off-schema, at claimed 200x faster and 400x cheaper than generative calls.
Rules for wiring it in
- Separate deciding from generating: most agent loops burn full-price frontier calls on multiple-choice questions like "which tool" or "should I retry".
- Set the confidence threshold before shipping, not after the first incident — the calibrated probability is the actual feature.
- Someone already rebuilt Claude Code's compaction on top of it, hitting 4,575 stars in five days.
A quoted article adds that in 2026 agents get assembled on existing surfaces, technical moats are gone, and the money moved to whoever packages the workflow first.
Related event: Decision-only model JEV sparks debate over speed and limits(2 posts)→
More from coding & agent
- Typed Evals: open-source framework for calibrated Jev-powered LLM evaluation — Charming_Group_2950 · 2026-09-21
- diagram-design skill for Claude Code hits 41.5k GitHub stars with editorial-quality diagrams — bibryam · 2026-09-21
- Qwen-Image 2.1 open-weights 7B model gets Day-0 ComfyUI support, runs in 30s on a laptop — Lexius2129 · 2026-09-21
- Dev builds bare-metal autonomous agent system after OpenAI refused to refund lost quota — epicskyes · 2026-09-21
- Open-source 'jev' tool automates computer use with no LLM, sub-second latency — _AustinCalvert_ · 2026-09-21
- Ex-Stripe Exec: Ready Your Business for Agent Buyers With Checkout, Payment Element and Invoices — jeff_weinstein · 2026-09-21