What is Jev? It is not a chatbot. It is the AI that decides.

Jev vs ChatGPT: ChatGPT writes, Jev decides

The fastest way to understand Jev is to put the same job in front of both kinds of model. A chat model is asked to produce something. Jev is asked to choose something. Most of the software people build around AI is actually the second kind of job, even when it gets wired to the first kind of model.

Neither column replaces the other. The next generation of AI apps will use both: a text model where words are the product, and a decision model everywhere else. The rest of this guide explains why the decision column is suddenly cheap enough to use constantly.

How Jev works: state in, typed decision out

Every Jev call has the same shape. Your software supplies the state, meaning whatever the model should look at: an email, a row of data, a page's DOM, the positions of enemies in a game. Then it asks one or more typed questions about that state. Jev fills in every answer in a single pass, each with a probability, and hands the result back to your code.

The community built around Jev uses three question types, and you will see them named in almost every project in our index:

• Choice: pick one option from a list you supply, for example which tool an agent should call next.
• Score: rate something on a scale you define, for example how well a landing page matches the ad that sent the visitor.
• Noul: a yes-or-no judgment with a probability, for example whether a message is spam.

Here is a simplified, illustrative exchange. The exact request schema is in TypeSafe's documentation; the point is the shape of the answer, not the field names.

state:    "Our payment page has been down for 20 minutes and I have a demo at 3pm."
question: priority -> choice of [urgent, normal, low]

answer:
{
  "priority": "urgent",
  "probability": 0.94
}

Because the answer can only ever be one of the options you wrote, it slots straight into an if-statement. There is no parsing of prose and no retry loop for malformed JSON. That is the main reason developers describe Jev as feeling more like a function than a chatbot.

Why Jev's speed and price change what software can do

A chat model has to generate its answer one token at a time. Jev skips that loop, because it scores a fixed set of answers instead of writing new ones. According to InfoWorld's launch coverage, TypeSafe reports end-to-end latency between 70 and 500 milliseconds on decision-shaped workloads, compared with several seconds for the chat models it tested.

Pricing follows the same logic. Coverage at launch reported a listed price of $0.042 per million input tokens and no charge for output, since there is almost no output to bill. Prices change, so treat that as the launch listing rather than a quote.

Put those together and a product can afford to make a decision on every click, every row and every frame. The clearest demonstration is TypeSafe's own Doom demo. As The Register reported, Jev never sees the screen; software turns the game into structured state and Jev answers one small question about ten times a second. According to the same report, TypeSafe put the cost at around $7 an hour and admitted that an ordinary scripted Doom bot would play better. That candour is the useful part: the demo shows speed, not superhuman play.

Ten decisions you make every day that Jev can make at scale

Jev use cases are easiest to spot when you stop thinking about AI as a writer. Look for places where a person, or an expensive model, is repeatedly choosing between a handful of known options:

1. Which email deserves a reply today?
2. Which task should happen first?
3. Which sales lead deserves attention?
4. Which clip is worth editing?
5. Which support ticket is urgent?
6. Which part of a web page is clutter?
7. Which customer needs a human instead of a bot?
8. Which paragraph of a long document actually matters?
9. Which tool should an AI agent call next?
10. Is this shell command safe to run?

Every item on that list already has at least one open source experiment behind it. Browse the directory, or start with our curated list of open source Jev projects, where each entry links to the exact file that calls Jev.

What Jev is not: limits worth knowing before you build

Jev does not write essays, draft emails or generate code. Projects that seem to do those things pair Jev with a text model: the text model writes, and Jev picks, ranks or checks. Browser agents built on Jev, for example, call a text model only when they need to type something into a form.

You will also see the line that Jev "never hallucinates". That claim deserves care. A model that can only return answers you defined cannot invent a citation or a tool name, which removes a whole class of failure. It can still confidently choose the wrong option. Calibrated probabilities help, because you can send low-confidence cases to a person or a bigger model, but they do not make the model infallible.

• Keep ordinary code as the final authority. Jev recommends; your program decides whether to act.
• Set thresholds from your own labeled data, not from launch benchmarks.
• Be cautious anywhere mistakes cost real money. The most viral trading demos include one whose author says it lost money.

Getting started with Jev

Jev itself is a hosted model, so you need API access from TypeSafe. The official SDKs are the simplest starting point:

1. TypeSafe Python SDKtypesafe-ai/typesafe-sdk-pythonOfficial TypeSafe Python SDK with synchronous and asynchronous clients for Jev System One, plus question and answer types.★ 202
2. TypeSafe JavaScript SDKtypesafe-ai/typesafe-sdk-jsThe JavaScript and TypeScript SDK published by TypeSafe, with typed Jev requests and answers.★ 226

If you would rather see it working first, the viral Jev demos we fact-checked show what people are building, and the Jev-like models category lists independent open alternatives you can run on your own hardware. For a guided walkthrough, our video tutorials collect the clearest explainers, and our analysis of how the press covered Jev separates the reporting from the hype.

Keep reading