Point at a canvas, ask for a shape, then tell it to move or change colour. Give a model the state of a driving simulator and let its decisions help steer the car. People are already building this with Jev, and I think it's seriously cool.
TypeSafe introduced Jev on September 15. What excites me is how many places you can put AI when small decisions become fast and cheap. Inside an editor. Inside a simulation. Inside the little interactions that make an app feel natural to use.
Look at what people are building
In jev-canvas, speech recognition and finger tracking let you create, move, resize and recolour shapes by talking and pointing. Jev selects the action, target and location; the app changes the canvas. That's an interface I'd love to build on.
JevPilot puts it in a driving simulator. Jev receives information about the road, traffic and possible paths, then chooses where to go and at what speed. The simulator handles the vehicle controls and collision braking.
As someone who builds interactive systems, this is the part that gets me. A lot of small decisions can add up to a completely different way of using software. I immediately start thinking about characters reacting to a trainee, or creative tools that understand what you're pointing at.
Small questions, usable answers
Jev is TypeSafe's first "System One" model. You supply information and specific questions; it returns decisions and probabilities that your code can act on.
There are three question types:
| Type | What you ask | What comes back |
|---|---|---|
| Choice | Which issue from this list? | One option, probabilities for the options, and confidence |
| Score | How urgent is this, using these described levels? | A probability-weighted score, the level probabilities, and confidence |
| Noul | Does this message request a replacement? | A probability from 0 to 1, with no separate confidence field |
Take a customer message: "The kettle arrived broken and you charged me twice. Can you send another one?" Separate questions can identify the damaged item, duplicate charge and replacement request. Code checks the order and sends each issue to the right place. The customer gets to explain the problem in their own words.
You can ask independent questions together, so those checks can happen in one call. Each question reads the same supplied information independently; dependent steps stay in code.
Cheap enough to use throughout an app
TypeSafe currently lists Jev 1.13 at $0.042 per million input tokens, with no output-token charge. It takes text, including structured text records; photos and audio need preprocessing.
As an illustration, 10,000 requests averaging 1,000 billable input tokens each would cost $0.42 in Jev input charges. That includes the questions and options in the token count; retrieval, other models and retries are additional costs.
That changes what I want to build. An inbox that picks out the school email needing a reply. An assistant that checks whether a draft actually answers the request. A tool that chooses a useful next action as you work. At that price, I want to put those decisions in far more places.
The pieces fit together
TypeSafe's smart-home demo shows the combination: Jev identifies the device and action, while an LLM handles conversation and splits compound requests. Each part gets a job it suits.
That's how I'd build with it: Jev for routine interpretation, code for rules and actions, a reasoning model for investigating difficult problems, and a generative model when someone needs a written reply.
There are useful measurements of those roles, too. LiteLLM's routing benchmark put Jev 1.13's median classification time at 126.81 ms against Claude Haiku 4.5's 688.40 ms: about 5.43 times as fast. It used 80 authored prompts, repeated three times, and measured routing decisions rather than the eventual answers.
LangChain used Jev to evaluate agent runs. On five weather-agent traces evaluated 100 times each, it matched the human pass/fail labels throughout and produced consistent scores. That gives builders a concrete evaluator pattern to work from; the five-example dataset defines the scope of the result.
What I'd keep in code
I'd keep arithmetic, date comparisons and permission checks in code. TypeSafe documents weaknesses with numbers, dates, indirect reasoning and misleading input. Jev chooses within the options you supply, which prevents invented options, but a wrong choice is still possible.
Its confidence field summarizes the probability distribution for Choice and Score. I'd use it to help route uncertain cases to another model or a person, with thresholds checked against real examples. It isn't an accuracy guarantee.
What I want to build
If you run a team: look at the small judgments people repeat all day. Which message needs attention? Who can resolve this request? Start with one useful workflow and measure the time saved and the corrections needed.
If you write code: build an interaction you'd enjoy using. Let someone point, speak, or describe what they want, then use Jev to choose among the actions your app supports. Compare responsiveness and errors with your current approach.
I want to use this in interactive tools and training simulations. Seeing what people have already built makes me want to get on with it.
Sources linked throughout. Checked September 22, 2026. Project descriptions come from their creators, pricing from TypeSafe, and benchmark results from the linked authors. My proposed uses and cost calculation are illustrative.