Jev doesn't chat. It returns a choice, a score, or a yes/no probability your code can branch on. Here's how that works and how to try it in five minutes.
If you've built on top of a large language model, you've probably written this code before: ask the model to "answer only with billing, bug, or account", then parse the reply, then handle the cases where it added a sentence, used a synonym, or broke the JSON. You needed one small branch, and you got a paragraph.
Jev, a model from TypeSafe AI, takes a different approach. It's not a chatbot, and it doesn't write text at all. You send it the state your application already has, plus a few typed questions. It sends back values your code can use straight away.
TypeSafe announced Jev on 15 September 2026 as its first public "System One" model. The name comes from the idea of fast, intuitive System 1 judgments (as opposed to slow, deliberate System 2 reasoning). These are the cheap decisions software makes all day: which queue does this ticket go to, how urgent is it, should this agent's tool call run? The hosted API opened to everyone on 21 September 2026, with no waitlist. The current public model string is jev-1.13.0, which is also available through the aliases jev-latest and jev-preview.
Jev is a closed, hosted model. TypeSafe hasn't released the weights, and TypeSafe says the current version isn't trained to generate text.
Every question you ask Jev is one of three types:
You can ask up to eight questions in one request. They're all evaluated against the same state, so the input cost is shared.
One detail matters for how you design things: Jev picks the branch, but your application still performs the action. Jev can say "billing". It won't issue the refund.
The sensible way to think about Jev is as a decision layer that sits between intent and action. Good starting points are narrow judgments that repeat:
What Jev isn't for: drafting, summarizing, explaining, coding, or long reasoning. For that you still want GPT, Claude, or another general model. In practice the pattern is hybrid. Jev decides the branch, and the LLM writes whatever a person will read. The independent guide explains this split in more detail on its page about what Jev is.
TypeSafe's published price is $0.042 per million input tokens, and output is free, because the response is a typed value rather than generated text. You can pay TypeSafe directly through its Console, or use the same list price on OpenRouter or Vercel AI Gateway.
There's also jevmodel.org, an independent guide and playground for Jev (it isn't affiliated with or operated by TypeSafe). It lets you try Jev in the browser without creating a vendor account:
The site is upfront that its per-token price is higher than going direct. You're paying for convenience, and at high volume TypeSafe or your existing gateway account is cheaper. For evaluating the model, the free tokens are usually more than enough.
Step 1: try it without code. Open the playground, paste in some text (for example "I was charged twice. Please fix this ASAP."), and add a choice question with a few labels. Check whether the output makes sense for your data before you write any code.
Step 2: create a key. Sign in, open the dashboard, create an API key, and store it on your server as JEVMODEL_API_KEY.
Step 3: send the request.
import os
import requests
response = requests.post(
"https://jevmodel.org/v1/systemone",
headers={"Authorization": f"Bearer {os.environ['JEVMODEL_API_KEY']}"},
json={
"model": "jev-latest",
"state": "I was charged twice. Please fix this ASAP.",
"questions": {
"category": {
"type": "choice",
"instructions": "What is this ticket about?",
"criteria": {"billing": "payments", "technical": "broken product", "other": "other"},
},
"escalate": {
"type": "noul",
"instructions": "Should a human review this now?",
},
},
},
timeout=15,
)
print(response.json())
A few rules from the docs are worth knowing: state can be a string or a JSON value up to 8,000 characters, each request takes 1 to 8 questions, and the endpoint allows 120 requests per minute per key. Add an Idempotency-Key header when you retry. The full walkthrough is in How to use Jev.
jev-1.13.0 once your thresholds are tuned, because aliases can move.
Jev is interesting because it's so narrow. It won't replace your LLM, but it can take over the fuzzy if statements you're currently paying a chat model to answer and then parsing. If you have a triage, routing, or guardrail step like that, it's worth an afternoon to test Jev on real examples.
Disclosure: jevmodel.org is an independent guide and isn't affiliated with, endorsed by, or operated by TypeSafe AI. Jev is a TypeSafe AI product. Facts in this post come from jevmodel.org and TypeSafe's public statements as of September 2026.