Jev, from TypeSafe AI: what it is, how to get access, and how I use it

Original author：tacosdedatos

Original source：https://www.youtube.com/embed/UMtP7i-ugW0

A few days ago TypeSafe AI released Jev, its first model. It doesn't chat, summarize or write code: it answers closed questions (choice, score, boolean) with a calibrated probability, in milliseconds, for almost nothing. This video covers what it is, why it's a big deal, how to get access (Vercel AI Gateway, OpenRouter, Cloudflare Workers AI and TypeSafe direct) and how I use it in hm, my terminal tool for checking text against my own taste. Chapters: 00:00 A model that can't talk 00:29 The problem: paying for prose to get one bit 00:57 Where Jev comes from (TypeSafe AI, Jevons) 01:21 What it is: a System One model 01:54 Three kinds of question: choice, score, boolean 02:36 The schema rules 03:00 Why it matters: latency and price 03:36 "Calibrated" means something specific 04:06 How to get access: four doors 05:18 How to use it: state + questions 05:54 My example: hm 06:48 Where it fails 07:41 The numbers, and the real question Links: - TypeSafe AI: https://typesafe.ai - Vercel AI Gateway (typesafe-ai/jev, AI SDK 7 evaluate): https://vercel.com/ai-gateway - OpenRouter (typesafe/jev-1.13, decisions endpoint, alpha): https://openrouter.ai - Cloudflare Workers AI (typesafe/jev): https://developers.cloudflare.com/workers-ai - Known weaknesses: docs.typesafe.ai/model-jaggedness - Evaluation: evals.typesafe.ai Latency, price and evaluation figures are TypeSafe's own; as of September 19, 2026 there is no independent reproduction. Free access on Vercel runs until September 25, 2026. This video was made by an AI agent: script, voice (my ElevenLabs clone), code-driven animation (Remotion) and mix.