Model Profile

What Is Jev by TypeSafe?

Jev is a model from TypeSafe AI built for a different job than most AI tools. It does not write text, chat, or code. It reads a state, such as a support ticket or an order record, answers a set of typed questions, and returns each answer with a probability. This guide explains what Jev does, how it is called, what it cannot do, and which claims are vendor statements rather than independent results.

By Stackari·10 min read·AI Models

The Short Answer

Jev is a decision model for software. Your application sends it data and a list of questions with defined answer types. It replies with a choice, a score, or a yes/no probability. Your code then acts on the answer.

TypeSafe calls Jev its flagship model and the first of a class it names “System One” models. Those models return judgments in a single pass rather than generated prose.

What Jev Does

Most AI tools generate text one piece at a time. Jev is built to skip that step. It takes messy real-world input, such as an email, a support ticket, or a game state, and answers questions you define in advance.

A simple real-world example

Imagine you run an online store and receive hundreds of customer messages every day. A customer writes:

“My order has not arrived, and it was supposed to arrive three days ago.”

Your software needs to decide which team should handle the message. Jev chooses from the options your software defines, such as Shipping or Billing. Your application then sends the message to the shipping team. Jev makes the decision, and your software performs the action.

Illustrative example, not a live Jev API result.

What does Jev output?

Choice: pick one option. Which team should get this message? Shipping (0.92 probability), Billing (0.08).
Score: rate something. How urgent is this request? 8/10. The scale is defined by the developer.
Noul: estimate yes or no. Does this request need a human to review it? Yes, with a probability of 0.95.

The numbers are hypothetical, not measured results or guarantees of correctness.

TypeSafe describes System One models as built for “fast, structured decisions that software can use directly.” The output is designed for code to read, not for a person to read.

A simple way to picture it

An ordinary if statement checks a fixed condition. Jev handles conditions that need judgment, such as “does this message sound urgent?” or “which department should handle this?”, and returns an answer your code can branch on.

The Three Question Types

According to TypeSafe’s documentation, Jev answers three kinds of question:

Choice

Picks one answer from a list you define, for example routing a ticket to “billing” or “technical.” Returns probabilities for each option.

Score

Places the state on an ordered scale you define, such as a 1 to 5 rubric. Returns a numeric position.

Noul

Answers a yes or no question with a probability from 0 to 1. For example, 0.95 means very likely yes.

All questions in one request are answered separately, using the same input. Some answers also include a confidence score, which is how sure Jev is about its own answer. TypeSafe’s documentation lists which question types include one.

How It Is Called

Developers can call Jev through a client SDK or a direct HTTP request to TypeSafe’s System One endpoint. The model field selects the version. jev-latest is used in TypeSafe’s examples and always points to the newest version. A fixed version such as Jev-1.13.0 is also available.

What goes in a request

State: text, JSON objects, or an array of text. Images, audio, and video are not supported.
Questions: a named set of typed questions, each using Choice, Score, or Noul.

What comes back

An answer for each question, labeled with the name you gave it, plus probabilities or a confidence score.
E.g Your application sends in a message or task, Jev returns a structured decision, and your application’s code uses that answer to route, filter, rank, or flag something. Jev does not open websites, send emails, or take actions just because it returned an answer.
No generated sentences and no explanation of reasoning.

Lock to one version in production. jev-latest is a moving target. When TypeSafe updates it, answers can shift, and any cutoff you set (for example, “act only if 90% sure”) may stop working as expected. Pick a fixed version, and retest before upgrading.

Pricing and Access

One listing reports input pricing of $0.042 per million tokens with output at $0 per million tokens. Hosting platforms may bill differently, and pricing can change, so check the current figures before budgeting.

Jev is available through TypeSafe’s own API and through third-party routes, including Vercel’s AI Gateway and OpenRouter. Coverage describes a move out of early access with the waitlist removed, though the exact date is reported inconsistently across sources.

Rate limits and context window size are listed on TypeSafe’s pricing page. The OpenRouter listing shows a 262K context length for Jev 1.13.

Do you need a subscription? None of the sources I reviewed mention one. The pricing described is pay-as-you-go: you pay per million input tokens you send, so a test costs only what you use. Confirm on TypeSafe’s pricing page, since terms can change.

Vendor Claims vs Confirmed Facts

Much of the public coverage repeats TypeSafe’s own figures. Treat the following as claims until you test them yourself.

Claim	Status
Latency of 70 to 500 ms end to end; most calls around 100 ms	Vendor benchmark. Measure on your own workload.
Cheaper than general-purpose language models on selected workflows	Vendor statement. Depends on your task and volume.
Probabilities match real outcomes on average	TypeSafe says this is measured across many answers together. It does not guarantee that any single answer is right.
“Can’t hallucinate” or “deterministic”	Not supported by the documentation or independent sources. Jev can still give wrong answers.
“200 times faster” than chatbots	Appears in a headline of one article without supporting detail in its text. Do not rely on it.

Limitations

Text input only. Images, audio, and video are not supported yet, according to the documentation.

Not a text or code model. Jev is not trained for generation, including writing code or prose.

Not a calculator. It does not count reliably and reads dates as text. Put arithmetic and date comparisons in your own code.

Accuracy depends on input size. Coverage reports that accuracy drops as irrelevant content grows, so keep inputs focused.

Scores are averages. A score of 90% means that, across many answers like it, about 9 in 10 are right. It does not guarantee that this one is. Build a review step for important decisions.

Who Should Consider It

A good fit

Routing or classifying incoming tickets, leads, or messages
Scoring records against a fixed rubric
Deciding which tool or next step an AI agent should take
Adding guardrails that check an agent’s output before it acts
Pulling out key details from messages, such as order numbers, dates, or the product a customer means
Flagging messages or orders that a person should look at, such as a complaint about safety
Checking whether product descriptions follow your rules, such as required fields or banned claims
Deciding which leads are worth a sales call, using your own scoring rubric

These are possible uses based on the three question types, not tested case studies. Try them on your own data first.

A poor fit

Drafting emails, articles, or summaries for people to read
Generating code
Anything that needs media input
Decisions that need a written explanation for an auditor

Test before you commit

Gather 100 or more real examples where you already know the right answer, such as past support tickets and where each one was sent. Run each one through Jev and through your current method. Compare accuracy, speed, and cost, then decide.

Can you test without a subscription? As far as the sources show, yes. Jev is priced per use, so a test costs only the tokens you send. You will still need access through TypeSafe’s API or a gateway that offers Jev. Check the terms before you start.

Sources

FAQ: Jev by TypeSafe

Is Jev a chatbot?+

No. It does not hold conversations or write replies. It answers typed questions about a state you send it.

What does “System One” mean?+

It is TypeSafe’s name for a class of models that return structured judgments with probabilities in a single pass, rather than generated text.

Can Jev write code?+

No. It is not trained for generation, including code.

Does Jev accept images?+

Not yet. Input is text, JSON, or arrays of text.

Is Jev accurate?+

Accuracy depends on your task. TypeSafe says its probabilities match real outcomes on average across many answers, not for each individual one. Test it on your own cases before relying on it.

How much does it cost?+

One listing shows $0.042 per million input tokens and $0 for output. No subscription is mentioned in the sources I reviewed; you pay per use. Pricing varies by platform and can change, so check current rates.

Should I use the latest version?+

For production, choose a fixed version. If you set a cutoff for how sure Jev must be, a newer version can change which answers pass it, so retest after any upgrade.