What Is Jev AI? TypeSafe AI's Decision-Making Model

On September 15, 2026, TypeSafe AI launched Jev AI - the first model in a product line the company calls the "System One Model." The biggest difference between Jev and ChatGPT, Claude, or Gemini is that it isn't built for conversation or content writing. Instead, Jev takes in situational data plus a structured question and returns a specific decision - for example, selecting one of several predefined options, scoring something on a scale, or answering yes/no along with a probability. That's why Jev has generated so much discussion in the developer community, particularly around its claims of near-elimination of "hallucination," a problem common to typical language models. In the article below, Aniday breaks down how Jev AI works, who it's for, and how to get started.

What Is Jev AI?

Jev is an AI model that TypeSafe describes as an "intelligence function call": unstructured state data goes in, and a clearly typed decision with an attached probability comes out. The "System One Model" concept draws on the fast-versus-slow thinking framework (System 1 and System 2) from psychologist Daniel Kahneman's book Thinking, Fast and Slow. TypeSafe argues that many software tasks - such as request routing, risk scoring, or content classification - don't need a model that "thinks slowly and writes long-winded prose," but rather need a decision that's fast, reliable, and in the right format.

Technically, Jev processes questions through a parallel sampling mechanism rather than generating text sequentially, word by word, the way large language models do. As a result, the entire response is produced almost simultaneously, significantly cutting latency. TypeSafe also emphasizes that because output must always match a pre-declared structure, the model "cannot" produce a choice outside the allowed list - this is the basis for the claim that Jev doesn't make type errors. However, some independent analyses, including one from The Register, note that the "hallucination-free" label needs to be read narrowly: Jev can still make a wrong decision - it's just that the decision will always come back in the correct, specified format.

Who Is Jev AI For?

• Developers building AI agents: for a fast routing step, determining which tool should be called next, or which requests need to be escalated to a human - Jev AI serves as a lightweight decision layer, separate from any text-generation component.

• Operations and risk teams: fraud scoring, prioritizing customer support tickets, or filtering policy-violating content all map well onto Jev's three decision types.
• Teams that need to verify AI output: Jev AI can act as a guardrail for the output of another language model, confirming whether it complies with a defined set of rules.

Notable Features of Jev AI

• Three decision types (Choice, Score, Noul): Choice selects one option from up to 255 possibilities; Score rates something on a scale of 2 to 10 levels and can return a decimal value; Noul returns a probability between 0 and 1 for a yes/no question.

• Processes multiple questions in parallel within a single call: since output tokens are free, you can ask several questions at once without incurring much extra cost or wait time.
• Low latency, priced by input tokens: responses in the 70-500 millisecond range, at $0.042 per million input tokens, with output tokens free.
• Always returns the exact declared structure: results come with a probability and confidence score for each option, so applications can handle downstream logic without parsing free-form text.
• Handles large volumes at low cost: TypeSafe cites real-world examples such as classifying more than 1,000 research papers for just $0.08, or a browser agent booking a flight in 7.1 seconds for $0.0039.

How to Use Jev AI

Since Jev is a developer-focused product, the main way to use it is by calling the SDK in code rather than through a web interface:

Step 1: Sign up for early access and get an API key at console.typesafe.ai/settings/keys (or via the Vercel AI gateway).

Step 2: Install the SDK.

For Python: pip install typesafe-sdk.

For JavaScript/TypeScript: npm install @typesafe-ai/sdk.

Step 3: Set your API key as an environment variable,

e.g. export TYPESAFE_API_KEY="sk-...".

Step 4: Write your question using one of the three decision types, along with the context data (state) Jev needs to make the decision.

Example in Python:

python

from typesafe_sdk import Choice, Noul, TypeSafeClient

client = TypeSafeClient()
response = client.system_one(
    state={"ticket": "I was charged twice"},
    questions={
        "department": Choice(
            instructions="Which department should handle this request",
            criteria={"billing": "Billing issue", "technical": "System error"}
        ),
        "urgent": Noul(instructions="Does the message indicate urgency")
    }
)

Step 5: Call the function and read the returned result, including the selected choice, the probability for each option, and an overall confidence score.

Step 6: Wire the result into your application logic
For example, automatically routing the request to the right department or deciding whether human intervention is needed.

Tips for Using Jev AI Effectively

• Ask multiple questions at once: since output is free, take advantage of parallel questioning instead of making repeated sequential calls.

• Set different confidence thresholds by risk level: a read-only action can tolerate a much lower confidence threshold than a financial transaction.
• Filter your data before feeding it to Jev AI: accuracy drops when the context contains a lot of irrelevant information, so retrieve only what's actually needed first.
• Pin the model version: if you've already tuned confidence thresholds for your system, specify an exact version (e.g., jev-1.13.0) rather than using the latest, since newer versions can change without notice.
• Don't use Jev for counting or math: per community best-practice guidance, Jev doesn't handle counting, arithmetic, or comparing dates as ordered quantities well.

Jev AI Pricing

Since Jev is still in early access, new users need to join the waitlist to get production access. TypeSafe hasn't announced a fixed monthly pricing plan the way consumer products do — instead, it charges directly based on input token volume.

Limitations to Keep in Mind

• Not a chat or content-writing tool: Jev doesn't generate text, summarize, or write code.

• Weak at counting and calculation: the model interprets dates as text rather than ordered quantities, making it unreliable for arithmetic.
• Accuracy drops with redundant data: the more irrelevant information present, the lower Jev's chances of deciding correctly.

• Takes instructions literally: descriptions that aren't clear or that contain contradictions can easily lead Jev to produce unexpected results.
• The "hallucination-free" claim needs proper context: according to The Register, this comparison isn't fully equivalent to free-text generation models — Jev only guarantees correct formatting, not that the decision itself is always accurate.
• Still in early access: new users must join a queue, and the model may still change before its official release.

Overall Assessment

Jev AI represents a different approach within the AI landscape: rather than trying to build an assistant that can write anything, TypeSafe chose to build a model specialized in decisions that are fast, structured, and verifiable. Given its published speed and cost figures, plus a founder with a background in developing ChatGPT, Jev is a notable product for teams building AI agents or automation systems that need a reliable decision-making layer.

That said, this is clearly a tool for developers rather than a product for general consumers, and many of its performance claims still need time to be independently verified. If you're looking for an assistant to chat with or write content, Jev isn't the right fit. But if your work involves building automation logic for software, it's a tool worth trying during its current early access period.