BlogAI
Jev, or the AI that does not write, it decides
How much of our AI is really decisions?
When teams plug an LLM into a product, it usually looks similar. There is a prompt describing the task, a list of allowed answers and a request for JSON. The model is to read a ticket and return “billing”, “technical” or “sales”. Or judge whether a customer is upset. Or check whether the command an agent wants to run in a terminal is safe.
In such a task nobody needs text. What is needed is a single label and information on how far it can be trusted. Yet we pay for a model that assembles its answer token by token, can write a poem about cats, and sometimes thinks for dozens of seconds over a question a human answers in one.
Jev, a model that TypeSafe AI presented on 15 September 2026, aims at that gap.
What Jev is
Jev is the first public model in a category its creators call System One. The name evokes Kahneman: System 1 is fast, intuitive judgment, System 2 is slow, step-by-step reasoning. Models with a reasoning mode are closer to System 2. Jev is meant to be System 1 for software.
How Jev differs from an LLM
In short: an LLM answers with text, Jev answers with a decision.
The shape of the answer. An LLM generates text, even when we ask for JSON. If the provider enforces a schema (structured outputs), the format is correct, but inside it is a word chosen by the model. Some APIs return token probabilities, but those are the probabilities of consecutive tokens, not of decisions. Jev returns a value from a type we defined ourselves: a choice from a list (up to 255 options), a rating on a described scale, or a probability from 0 to 1. It has no way to return something outside the list.
Confidence. For a choice from a list and a rating on a scale we get a probability distribution over all options and a confidence level, and for a yes-or-no question just the probability. The model is trained with a method TypeSafe calls RLCD (Reinforcement Learning for Calibrated Decisions). Its goal is “honest” probabilities: if the model says 0.9, it should be right roughly nine times out of ten. Language models, when asked about their confidence, tend to be overconfident.
How it computes. An LLM writes its answer token by token. According to its creators, Jev generates all answers in a single pass, in parallel, so several questions in one request do not mean several times longer waiting.
Time and cost. TypeSafe states 70 to 500 ms per answer, against 3 to 329 seconds for leading LLMs. The price is 0.042 dollars per million input tokens, and the output is free. For comparison, Claude Haiku 5.5, Anthropic's small model, costs 0.10 dollars per million input tokens. The company itself admits it cannot prove that its price is not subsidized.
What it does not do. It does not write text, generate code, explain its decisions or choose the next step by itself. The documentation says plainly that it is a model for building software with AI, not for building agents.
What it looks like in practice
We send the API a state, that is, text or an object with data (for example the content of a ticket with metadata), and a list of questions. Each question has its own type:
choice: “which team should handle this?”, with the options billing, technical and sales and a short description of each;
rating: “how frustrated is the customer?”, on a scale described in words;
probability: “is the message urgent?”, with an answer from 0 to 1.
Each question returns a value and its confidence. More interesting than the API itself is how TypeSafe advises using it:
Ask narrow, single questions. A broad question hides several judgments under one answer.
Let the code do something different for a confident and an uncertain answer. Uncertain cases go to a human or to a more expensive model that can reason.
Set thresholds on your own data. Compare the model's confidence with its accuracy on your cases, and only then decide at what confidence you trust the answer.
This is classic machine learning, not a novelty of the LLM era. And it is probably the most valuable thing Jev brings: a discipline that ML teams have known for years, in a tool used like a prompt.
Where it makes sense
The common denominator is simple: many repeatable decisions, an answer from a closed list, and time that matters, because a user is waiting or an agent is executing hundreds of steps. Typical examples are triage of tickets and emails, spam and phishing filtering, and document review.
How we would approach it
We would not replace the LLM across the whole system. We would look in it for the steps that are decisions. On a few hundred real cases we would check accuracy, cost and time, and then set a confidence threshold: above it Jev decides, below it a reasoning model or a human. The creators do not publish a list of supported languages, so a test on Polish data is a condition, not a formality.
In the end
Jev will not replace language models, because it cannot do what they are best at: writing, translating, reasoning and conversation. But it can take over work they were never needed for. And it reminds us that not every question requires a conversation.
Read next
When classical ML beats an LLM
When someone says “let's do it with AI”, the answer more and more often is “let's send it to an LLM”. Yet a large share of business…
ReadFrom the Trenches to the Watchtower: How to Work as a Developer When an AI Agent Writes the Code
An AI agent can write in two hours as much code as a human writes in a week. But who is going to review all of it? At JDD 2026 in Krakow,…
ReadPrefer to see this in a shipped project rather than in theory? Have a look at our case studies