aimodelsautomationproductivity

Jev: the AI that does not chat, it decides

What TypeSafe Jev is, how it picks between options with probability, and where to try it: waitlist, Vercel and OpenRouter.

Read in Portuguese →

Camila runs support for a small store. Three people, around 200 messages a day: defects, exchanges, billing, simple questions. She wants the system to sort it alone.

The first attempt looks obvious: send each message to ChatGPT and ask for a summary of what to do. For 5 messages it works. For 200 it breaks in three concrete places. Camila still has to read the summary and decide where each message goes, because the text comes back free form and her system only understands one exact field. Each answer takes seconds, and the queue moves in milliseconds. And once the AI invented the category "VIP urgent", which does not exist in the database, and the message got lost.

So the need changes. She does not need a writer inside her code. She needs a bouncer. Someone who looks at each message and answers short: this is an exchange, 94% sure. Route it.

200 store messagesdefect / exchange / billingmessy inputJEVpicks between options70 to 500msexchange, 94%typed decision + probabilityroutes itself

You define the options, Jev returns the decision with probability

That is what Jev is for, from TypeSafe AI. Company of Diogo Almeida, one of the researchers behind ChatGPT. Camila defines the options up front, Jev takes the message and returns only the decision, with calibrated probability. No paragraph. Code can act on it.

Now one problem remains: what when it labels wrong? This is where probability becomes the operating rule. At 94%, the system routes it. At 57%, it sends it to Camila for review. Same logic to triage tickets, set a sales route, or check if another AI promised a discount that is not in the database. The check stays cheap because output is free and input costs $0.042 per million tokens, with answers in 70 to 500 milliseconds.

In practice, an engineer at Vercel swapped a ChatGPT Luna 5.6 classifier for Jev and got 5 to 18 times faster, with better accuracy. In another email test, Gemini was slightly more accurate but cost 10 to 20 times more. TypeSafe calls this a System One Model: built for fast decisions, not chat.

Real limit: you write the options. Poorly written options turn into confident errors. And part of the published tests use another AI answer as reference, not final proof.

Where to try

  1. Official waitlist with early access and TypeSafe site. Full announcement: Introducing System One Models and Jev
  2. Docs to see how to define options and read output
  3. Jev 1.13 on OpenRouter, to call via API today
  4. Vercel AI Gateway, to try at no cost for now

Start with 1 to join the line and use 3 to test today.

Sources

  • TypeSafe post 09/18/2026: https://typesafe.ai/blog/introducing-system-one-models-and-jev
  • TechCrunch 09/18/2026: https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/
  • Olhar Digital 09/17/2026: https://olhardigital.com.br/2026/09/17/inteligencia-artificial/conheca-o-jev-nova-ia-do-cocriador-do-chatgpt-que-nao-conversa-e-foi-criada-para-tomar-decisoes-dentro-de-softwares/
  • Docs: https://docs.typesafe.ai/