Jev does not generate text. TypeSafe describes it as a System One model: it reads the input and answers fixed questions with probabilities. You cannot chat with it, and TypeSafe says it is not built for producing text.

What does noul mean in Jev?

noul is the name of Jev's yes/no question type. Its answer is one number from 0 to 1, the probability that the answer is yes. Your code decides the cut-off, for example 0.5 for a plain yes or no, or higher when a wrong yes is costly.

Which Jev version is current?

Jev 1.13, with the pinned id jev-1.13.0. TypeSafe's aliases jev-latest and jev-preview both point to it as of October 2026.

What is Jev? TypeSafe's typed decision model, explained

Q: Does TypeSafe train Jev on my requests?

No. TypeSafe states that Jev is not trained on customer requests or responses. Zero data retention is available to enterprise customers.

Jev in one paragraph

Jev is a decision model sold as an API by TypeSafe AI. A request carries a state (a string, a JSON object or an array of text) and a set of questions, each with a type. Jev answers every question with probabilities: which option fits, where the state sits on a scale, or how likely a yes is. Because the answers are typed values, your code can branch on them directly. TypeSafe calls this kind of model a System One model. The current release is Jev 1.13.

How a Jev request works

Every call goes to one endpoint, POST https://api.typesafe.ai/v1/systemone, with a bearer API key. The body has three fields, all required:

state: the input to decide on, as a string, an object or an array.
model: which Jev to use, such as jev-latest or the pinned jev-1.13.0.
questions: a map from your own question ids to typed questions. Each question has a type, instructions and, depending on the type, criteria.

The response returns the model that answered, one answer per question id, and token usage. There is a second endpoint, GET /v1/models, that lists the models. The Jev API page goes through every field.

The three answer types

Type	API name	What it returns	Typical use
Choice	`choice`	One option from up to 255, with a probability for every option and a confidence from 0 to 1.	Routing, intent, classification
Score	`score`	A place on 2 to 10 ordered levels: the expected level, the probabilities and a confidence.	Urgency, risk, lead quality, review priority
Yes/No	`noul`	One probability from 0 to 1 that the answer is yes. Your code picks the cut-off.	Gates before an action, policy checks

Limits

Context: 64k tokens a request for the state and all questions together, and 32k for the state plus the longest question.
Input: Text only: a string, a JSON object or an array of text values. Images, audio and video have to be turned into text before Jev can use them.
Languages: English first; other languages are handled, but TypeSafe says not equally well.
Rate limits: 100,000 tokens and 80 requests a second per key. TypeSafe says the limits can change without notice, and offers higher ones on custom and enterprise plans.

Price

Jev costs $0.042 per million input tokens ($42 per billion). Output tokens are free. TypeSafe's documentation lists no free tier. At 300 input tokens a decision, a million decisions cost $12.60. The Jev pricing page works through other request sizes.

Where Jev is weak, by TypeSafe's own account

TypeSafe publishes a list of the situations where Jev 1.13 is less reliable (model jaggedness, reviewed 2026-10-02). In our words:

Literal reading: It answers the exact wording of a question and can miss the intent a person would infer.
Counting and numbers: Counts, numeric values and magnitudes are unreliable; TypeSafe advises doing them in code.
Dates and times: Dates are read as text, so ordering, durations and window checks are unreliable.
Indirection: Double negatives and questions that need several reasoning hops lose accuracy.
Large, noisy state: Unrelated content in the input lowers accuracy, so TypeSafe advises filtering it first.
Adversarial content: Text written to steer the model can change its answer.
Contradictory criteria: When the instructions and the criteria disagree, answers get worse.
Option order: It tends to favor the option listed first, so reordering options can change answers.
Text generation: It is not built to write text, and chaining choices to fake it is slow and weak.

Adversarial content is the one where the difference to DecisionNode is measured: on the prompt injection suite, Jev scores 0.600 and DecisionNode MoE 0.764 (Jev benchmarks, preliminary, October 2026).

How Jev handles your data

TypeSafe states that Jev is not trained on customer requests or responses. Zero data retention is offered to enterprise customers. This is a point where Jev's defaults are stricter than DecisionNode's: DecisionNode uses requests for training unless you opt out, and offers zero retention on every plan.

When Jev is a good fit

Your inputs are English text or JSON, and your answers are choices, scores or yes/no.
The decisions lean on general world knowledge, where Jev scores 0.820 against DecisionNode MoE's 0.698.
You want official Python or JavaScript SDKs.
Your requests must never be used for training, by default.

When to look at DecisionNode instead

Your inputs include images or video: DecisionNode reads up to 16 images a request and video as frames.
You need numbers, points or boxes as answers, not only choices, scores and yes/no.
The state is written by users, web pages or other models that may try to steer the answer.
You run large jobs that can wait: batch jobs cost 30% less on MoE and 50% less on Flash.
You want to start without paying: 50 DecisionNode Flash requests, 20 DecisionNode MoE requests and 10 booster calls a day.

The Jev vs DecisionNode page compares the two in full, suite by suite.

Sources: TypeSafe's introduction, models and API pages, checked October 12, 2026.

What is Jev?