Jev: a model that decides instead of writing

Most of what we call AI today is built to write. You send in a question, and a model produces an answer one token at a time. But a surprising amount of what software needs from a model isn't writing at all. It needs a decision: which category does this belong to, how urgent is it, is it safe, is it hard. Jev, a new model from TypeSafe AI, launched on September 15, 2026, is built for exactly that, and it writes nothing.

What Jev does

You give Jev some text and a handful of typed questions, and it answers all of them at once. The questions come in three shapes. A choice question asks it to pick from a list, for example whether a message is small talk, a simple question, code or a reasoning problem. A score question asks it to place something on an ordered scale, such as how difficult a task is from 1 to 10. A yes/no question returns the probability that the answer is yes, for instance whether a text contains personal data or passwords.

Every answer arrives with a probability and a confidence value, and it always comes back as well-formed, typed data your code can use directly. There is no prose to parse and no JSON to hope for. Jev can tell you it is 92% sure a request is simple, or that it is torn between two categories, and that second kind of answer is often the more valuable one.

Where "System One" comes from

TypeSafe calls Jev a System One model, and the name borrows from psychology. The idea of two modes of thinking was popularized by Daniel Kahneman in Thinking, Fast and Slow, building on earlier work by the psychologists Keith Stanovich and Richard West. System 1 is the fast mode: automatic, intuitive, almost effortless. It's what lets you recognize an angry face, finish the phrase "bread and ...", or sense at a glance that a question is an easy one. System 2 is the slow mode: deliberate, step by step, and tiring. It's what you use to multiply 17 by 24 or work through a difficult problem.

Seen this way, the language models most people talk about are System 2 machines. They build an answer word by word, and reasoning models have long spent extra time thinking out loud before they reply. That is powerful, but the cost in time and money is paid on every request, including the ones that never needed it.

A System One model gives that up on purpose. Jev generates no free text at all, so it never waits for the next word. All of its answers are produced in parallel, the way a gut reaction arrives whole instead of as a sentence. Asking it three questions takes about as long as asking it one.

Fast, cheap and accurate

That design is what makes Jev cheap. TypeSafe quotes 70 to 500 milliseconds end to end and a price of $0.042 per million input tokens, with free output.

In TypeSafe's own workflow evals, which cover four workflows, Jev scores 67.8% accuracy, level with Claude Sonnet 5 and not far behind Opus 5 at 73.1%, with Haiku 4.5 at 53.6%. The difference is in cost and speed: about 0.4 seconds and $0.0004 per case for Jev, against 37.8 seconds and $0.18 for Opus 5.

Note that these are TypeSafe's own benchmarks, not tests we have run ourselves.

Jev plays Flappy Bird

Getting Jev to categorize messages, images and the like is genuinely useful, and honestly pretty cool. But we figured Flappy Bird would be more fun. So we built a Flappy Bird game split in two, with Jev playing on the left and GPT Luna, OpenAI's smallest and fastest model, playing the same game on the right.

Flappy Bird suits a System One model because the whole game is one quick decision repeated over and over: flap or don't flap. The demo keeps count of how many decisions each side has made and how many it makes per second, which is where the difference between the two shows up. In the recording Jev is making roughly 3.8 decisions a second against roughly 0.7 for GPT Luna. A model that is a little slower to answer is a little late to every flap, and in this game being late means hitting a pipe. Watching the two birds side by side makes the point about latency better than any table does.

What Jev is not

Jev can't write, explain or reason across several steps, and it isn't meant to replace a language model. Its natural place is in front of one, as the quick judgment that decides what should happen next. It can also be wrong. A confident answer can still be incorrect, so the confidence values need to be interpreted and tested against real data rather than trusted blindly. And because checking whether something is sensitive means sending it to the classifier first, privacy-conscious teams will want to look at self-hosted options as well.

Where it fits

Relevant uses are the ones TypeSafe points to: ticket triage, intent detection, moderation, scoring, safety checks, and routing a request to the right model depending on how demanding it is. That last one can bring lower costs and higher efficiency when easy requests stop being sent to the most expensive model, while quality can stay at about the same level.

How we use it

We already use Jev in projects we deliver and in our own products, for example to categorize requests, and also to route requests to the right model based on how complex the prompt is. That keeps costs down for our customers and gets them faster answers. We also use it for product classification, where the same fast, typed answers help us, for example, sort products into the right categories.