What Is Jev? What an AI That Returns Decisions Instead of Prose Can Do

September 19, 2026Updated: September 19, 2026Articles

TypeSafe AI released Jev in early access on September 15, 2026, as its first “System One Model.” Jev is not an AI that writes prose like ChatGPT or Claude. It specializes in decisions: choosing one of several predefined options, assigning a rating, or returning the probability that a condition is true.

In a post by Niku that a reader shared with me, you can find 30 examples involving browser control, games, code review, and more. Videos of Jev playing Mario or Doom are eye-catching, but its potential for work becomes clearer if you focus on situations where a person reads some text and chooses the next step from a fixed set of options.

Jev returns decisions in a predefined format instead of writing prose

When you ask a conventional generative AI a question, it returns a free-form response. That flexibility is useful, but it can also add unnecessary explanations or produce wording you did not specify when all you need is for the model to choose “sales,” “support,” or “other.”

You give Jev the text it should evaluate, a question, and a set of possible answers. It returns the selected answer, probabilities for each option, and a confidence value. TypeSafe AI describes this design as “unstructured state in, typed probabilistic decisions out.”

Comparison	ChatGPT, Claude, and similar models	Jev
Primary output	Prose, code, images, and more	Predefined choices, scores, and true-or-false probabilities
Best suited for	Explanation, summarization, creation, and conversation	Classification, scoring, branching, and verification
Output flexibility	High	Limited to a range defined in advance

“System One Model” is TypeSafe AI’s product term. Jev is not intended to replace a general-purpose chatbot. It is easier to understand as a model whose role is narrowed to making rapid, repeated decisions.

Three ways to ask: Choice, Score, and Noul

Jev’s API provides three main ways to ask questions.

Choice: Returns one item from a fixed set, such as sorting an inquiry into “billing,” “technical support,” or “cancellation.”
Score: Assigns an ordered rating, such as evaluating the severity of a bug on a scale from 1 to 5.
Noul: Returns the probability that a condition is true, such as whether an issue requires an urgent response.

Choice and Score return confidence information alongside the result. You could automatically route a request when confidence is high and send it to a person for review when confidence is low. Confidence does not guarantee that an answer is correct, however. You need to test the model on your own data before deciding when an action can be automated.

What Jev can do

Classify inquiries and spreadsheet rows

Jev can read emails, comments, and support requests and assign a category that matches their content. It can also label rows in a large dataset as “included,” “excluded,” or “needs review.” Tasks in which people repeatedly choose from the same set of options are a practical place to test Jev.

Score importance and risk

Jev can evaluate the quality of a post, the severity of a bug, or the risk of an action according to criteria you define in advance. Because it can consider meaning that simple keyword searches may miss, it can help determine what should be reviewed first.

Choose the next action for an AI agent

When an AI agent can use several tools, Jev can choose whether it should search, calculate, ask a person a question, or stop. Vercel also offers Jev through AI Gateway and lists selecting the next tool or subagent—and deciding whether to continue, retry, ask, or stop—as example use cases.

A system could use a generative AI to write prose while assigning only branching decisions to Jev. If you are interested in designs that combine multiple AI components, I also covered them in What Is Graph Engineering? Read This Before Adding More AI Agents.

Read state and select an action at short intervals

In published game and browser-control prototypes, Jev receives the current state as text and selects the next key to press or place to click. This approach is suited to real-time processing, where long waits between actions are impractical.

One important limitation is that Jev does not directly understand screens, video, or audio. Its input is text. Other parts of the surrounding system must organize on-screen elements or transcribe audio. A more accurate description is not “an AI that watches a game and plays it,” but “an AI that reads an organized description of the current state and selects an action from predefined options.”

The speed and price are appealing, but the figures come from the provider

TypeSafe AI lists Jev’s response time as 70 to 500 milliseconds and its price as $0.042 per million input tokens. There is no output charge. A single API call can evaluate multiple questions in parallel, so the same text could be checked for category, urgency, and risk at once.

The company has also published comparisons in which Jev is faster and less expensive than existing large language models. These evaluations were created by TypeSafe AI itself. Actual speed and accuracy will vary with the number of questions, input length, network conditions, and the data being evaluated. Before adopting Jev, you should measure it with data that resembles your own use case.

Limitations to understand before trying Jev

It cannot write prose or code: Pair it with another generative AI when you need explanations, replies, or summaries.
It accepts text only: Images, audio, and video must first be converted into text or structured data by another system.
A correctly formatted output can still be wrong: Even though Jev will not return prose outside the supplied options, it can misclassify or mis-score an item.
English is its strongest language: The official documentation says it can handle languages including Japanese, but they may not reach the same level of accuracy as English.
It is currently in early access: The model and specifications may change. TypeSafe AI currently offers a waitlist, and Jev is also available through Vercel AI Gateway.

TypeSafe AI uses the phrase “zero hallucinations” to describe outputs that cannot fall outside a defined schema. That does not mean every decision is correct. Actions with significant consequences—such as sending, deleting, or purchasing—should not be executed solely on Jev’s result. The system should involve a person when confidence does not meet an appropriate threshold.

I am more interested in small everyday decisions than game demos

Game-control demos make Jev easy to show. For practical work, however, I am more interested in its ability to make small decisions rapidly and repeatedly: choosing the responsible department, prioritizing items for review, or selecting the next tool to use.

A generative AI that writes and Jev, which handles decisions, could each take on the work they do best. If I were testing Jev for the first time, I would not immediately let it perform actions automatically. I would begin by asking it to categorize past inquiries and compare its choices with human decisions. Based on those results, I could then consider automatically routing only the cases where its confidence is high.

Summary

Jev is an AI that returns choices, scores, and true-or-false probabilities instead of writing prose. It can classify inquiries, assess importance, and choose an AI agent’s next action.

Its input is limited to text, and its decisions are not guaranteed to be correct. I want to look for tasks in my everyday work where I repeatedly choose from a fixed set of options, then explore how Jev might fit those tasks.