On 15 September 2026, TypeSafe AI released a model called Jev. It can’t write an email, a summary, or a line of code. It answers questions you define in advance, with a probability on each answer, in about a quarter of a second.
How it works
You give it a piece of text and a list of questions whose possible answers you write yourself.
The text: “My invoice lists two seats, but only one of us can sign in, and the login page keeps timing out.”
The question: which team handles this, billing, technical, sales, or other?
The answer: technical, at 96% probability, in 586 milliseconds. That test is from jevaiguide.com, an independent site, on 19 September.
The answer can only ever be one of the four options you listed. Nothing to parse, and no chance of a fifth team appearing out of nowhere.
There are three question shapes: yes or no, pick one from up to 255 options, and rate on a scale you describe. You can ask 20 questions about the same text in one request, and per those same tests it takes no longer than asking one.
Price and speed
$0.042 per million input tokens. Output tokens are free. A short request works out to about $0.00002.
TypeSafe states 70 to 500 milliseconds per call, and independent testing reports a median near 284ms. Their homepage claims 193.6x faster and 444.6x cheaper than frontier LLMs, from evaluations they designed and ran themselves.
What it’s good for
The pattern fits decisions a person could make at a glance, where you already know the full list of possible answers.
Classification and routing. Which team handles this ticket, which category a document belongs to, which of your tools should answer a request. You write the options, the model picks one, your code branches on it.
Filters and guardrails. The same thing with two options. Is this spam. Is this abusive. Is this someone trying to jailbreak the model behind it.
Scoring. Anything with levels you can describe: urgency, severity, how closely a candidate matches a role. The answer can land between two levels, which helps when you want a threshold rather than a label.
Real-time interface work. Anything that reacts while someone is still typing, scrolling, or playing needs an answer in well under a second. At 4 seconds a round trip, you build a different product.
Checking another model’s output. Take what an LLM wrote and ask a list of yes or no questions about whether each claim appears in the source.
Volume. At $0.042 per million input tokens, classifying a million short documents costs roughly $17, assuming 400 tokens each. That’s my arithmetic on their published price, not a figure they quote. It changes which jobs are worth doing at all.
Compared to Claude or ChatGPT
| Claude, ChatGPT, Gemini | Jev | |
|---|---|---|
| What you get back | Text a person reads | One of the answers you listed, with probabilities |
| Built for | Writing, reasoning, conversation | Deciding, classifying, routing, scoring |
| Speed | Seconds to minutes | 70 to 500ms, per TypeSafe |
| Output cost | Several times the input price | Free |
| Can it answer outside your options | Yes | No |
| Can it be wrong | Yes | Yes |
TypeSafe says Jev “can’t hallucinate”, and in a narrow sense that’s guaranteed: it can only return a value you defined, so it can’t invent a category you never wrote. It can still pick the wrong one and report 99% confidence while doing it. Staying inside the rules and being correct are different properties.
Known weak spots for the current version: arithmetic, counting, comparing dates, multi-step reasoning, and long inputs padded with irrelevant detail. Text only, no images or audio. English is the most accurate language.
My read
In a typical LLM agent, the routing logic lives inside a long prompt written in English. “If the user seems angry, escalate.” That logic is hard to test, hard to review, and it shifts every time someone rewords a sentence.
With this shape, the branching sits in your code, where you can read it and put a threshold on it. The model supplies one judgment: how confident, out of these options.
Treat that as a hypothesis. The public evidence is a few weeks old, and most of the impressive demos are community posts rather than audited results. A confidence score is still a claim, and 94% at scale means a real number of wrong calls. If being wrong is expensive, you still want a second check.
Sources
- TypeSafe AI, “Introducing System One Models & Jev”
- jevaiguide.com, independent documentation, including the response-time tests quoted above
- typesafeai.app, an independent directory of public Jev demonstrations
- “Jev Explained: Demos and Use Cases”, Syntax, 17 September 2026