What Is Jev? TypeSafe AI’s System One Model Explained for Beginners

If you’ve come across the name “Jev” in an AI agent’s config file, a guardrail library, or a recent write-up about building safer coding agents, you’re not imagining things – it’s a real product from TypeSafe AI, and it’s spreading fast because it solves a problem large language models are genuinely bad at: making a fast, reliable, narrow decision without any risk of making something up. So, what is Jev, exactly? That’s what this post walks through.

Jev Isn’t a Chatbot

Every model you’ve probably used – Claude, GPT, Gemini, or a local model like Qwen or Llama – is a text generator. You give it words, it predicts the next ones, and it keeps going until it has produced a paragraph, a function, or an essay. Jev doesn’t do any of that. TypeSafe AI calls it “System One” – a nod to Daniel Kahneman’s idea of fast, intuitive thinking, as opposed to the slower, more deliberate process of generating a full response – and it’s built to do exactly one thing: look at a situation and pick from a small, predefined set of answers. It’s closer to a very well-calibrated classifier than to a language model, even though it’s trained the way modern AI models are.

That distinction matters more than it sounds. A text generator can always keep talking – it can invent a plausible-sounding justification for anything, including a wrong answer. A model that can only choose between “yes” and “no,” or between three named options you supplied, has nowhere to hide. It either picks one of your options or it doesn’t respond at all.

Typed Questions In, Typed Decisions Out

Using Jev looks less like prompting and more like calling a function. You send it two things: some unstructured “state” (plain text describing the situation) and a list of typed questions. Question types include things like Choice (pick one of N labelled options), Score (a numeric rating in a range), and Noul (a yes/no-style boolean). Jev replies with a typed answer for each question, plus a confidence value, in well under a second.

from typesafe_sdk import TypeSafeClient

client = TypeSafeClient(api_key="YOUR_TYPESAFE_API_KEY")

result = client.evaluate(
    state="User message: 'Can I get a refund for an order from 400 days ago?'",
    questions=[
        {"type": "choice", "id": "route", "options": ["refunds_team", "general_support", "escalate"]},
        {"type": "score", "id": "urgency", "min": 0, "max": 10},
    ],
)

print(result["route"]["answer"], result["route"]["confidence"])
print(result["urgency"]["answer"])

There’s no prompt engineering here, no system message, no worrying about the model going off-script mid-sentence. You define the question shape once, and every answer comes back in that shape, every time.

Why It Can’t Hallucinate a Decision

“Hallucination” is really a side effect of open-ended generation: a model that’s free to produce any sequence of tokens will occasionally produce a confident, fluent, wrong one. Jev sidesteps the whole category of problem by never generating free text in the first place. Every answer has to be one of the options you defined. If none of the options genuinely fit, a well-calibrated confidence score will tell you that, rather than the model quietly picking one and sounding sure about it.

Speed and Cost

• Latency: typically 70-500ms per call, since there’s no token-by-token generation to wait for.
• Throughput: TypeSafe AI advertises 40-200x faster than calling a frontier LLM for the same kind of decision, at similar levels of accuracy on the decisions it’s built for.
• Cost: around $0.042 per million input tokens, with output free, since there’s no generated text to pay for.

That combination is what makes it practical to call Jev on every single step of an agent loop, rather than reserving LLM calls for the expensive, occasional check.

Where People Are Actually Using It

The most common pattern right now is guardrails for AI coding agents – gating a command before it runs and blocking anything that looks destructive, which we cover in detail later in this series. Beyond that, people are using Jev for support-ticket routing, content moderation triage, and as the decision layer in “computer use” agents, where something has to pick the next click or keystroke on every single step of a loop. It’s also available through several gateways – Vercel’s AI Gateway, Cloudflare AI, and OpenRouter – if you’d rather not manage a direct API key yet.

Where the Name Comes From

Jev is built by TypeSafe AI, a San Francisco company founded in 2024, and went into limited early access on 15 September 2026 alongside a $40 million seed round led by DCVC. The name itself is a small in-joke: it’s named after William Stanley Jevons, the 19th-century economist behind the Jevons paradox – the observation that making a resource cheaper to use tends to increase total consumption of it, not reduce it. The idea is that cheaper, faster machine decisions won’t just replace a few expensive LLM calls; they’ll get used in far more places than expensive ones ever were.

Where to Go Next

This is the first post in a short beginner series on Jev. Next up: making your actual first API call and understanding the response shape in detail. If you’re already running Claude Code against a local model, it’s also worth reading how Jev fits alongside a local LLM setup – they solve different problems and work well together.