TypeSafe AI’s Jev: the System One model that answers in JSON and tells you how sure it is about each decision

TypeSafe AI’s Jev is a model built to be used by software: it takes a structured question and always returns a decision in JSON, with a probability for each option and a confidence score from 0 to 1. It costs $0.042 per million input tokens, output is free, and it answers in 70-500 milliseconds. The confidence score is the reason I’m writing about it: TypeSafe AI presents it as a way to fight the overconfidence of large language models.

What is TypeSafe AI’s Jev?

Jev is the first model in the family TypeSafe AI calls System One models. The startup announced it on September 15, 2026, in a post signed by Diogo Almeida, TypeSafe AI’s founder and a former OpenAI researcher who worked on the training methods behind ChatGPT.

An LLM like ChatGPT or Claude generates text for a person to read. Jev generates decisions for a program to use. Jev doesn’t chat, explain or write code: it picks, it rates, and it tells you whether a statement is true.

What are Jev’s three primitives?

Jev works with three question types:

• Choice: picks an option from a list and returns the probability of each one.
• Score: rates the data you pass in on a scale of levels.
• Noul: tells you whether a statement is true, with a value from 0 to 1.

A concrete example: you hand Jev a support ticket and ask which team should get it, how severe it is and how urgent it is. You get three structured answers, ready to use in your code.

Why is Jev built for software?

Jev always returns structured data in JSON, in the shape you defined. That’s why Jev plugs straight into an application, with no text to interpret.

With LLMs, the JSON sometimes comes back with missing fields, and one missing field is enough to make the software reading it unstable. I’ve already written about why valid JSON isn’t enough to enforce your software’s constraints. LLMs are also slower: TypeSafe AI puts frontier models at 3 to 329 seconds, against Jev’s 70-500 milliseconds.

How much does Jev cost compared to an LLM?

According to TypeSafe AI:

• Jev costs $0.042 per million input tokens, or $42 per billion;
• output is free;
• existing LLMs charge $0.20 to $10 per million input tokens, with output at about 5 times the input price;
• in TypeSafe AI’s tests on workflows suited to a System One model, Jev came out 193.6 times faster and 444.6 times cheaper.

These are the company’s own figures, on tasks picked for Jev, and TypeSafe AI itself places them at the higher end of real-world gains. Measure them on your own use case before relying on them.

What is Jev’s confidence score and why does it matter?

The confidence score is the most interesting part of Jev. With Choice and Score, Jev returns the answer, the probability of each option and a confidence score that sums up how clear-cut the decision is. The confidence score is 1 when all the probability sits on a single option and 0 when it’s spread evenly.

The confidence score is a number, so it’s measurable. And a number is something you can put in a condition in your code. For example:

• at 0.9 you accept the decision automatically;
• at 0.6 you run an extra check;
• at 0.4 you hand the request to another large language model or to a person.

TypeSafe AI’s documentation suggests the same logic: higher thresholds for risky actions, lower ones for read-only actions, tuned on your own data.

How can Jev reduce LLM overconfidence?

One of the problems with today’s LLMs is that they sometimes answer "yes, I’m sure" when they aren’t. TypeSafe AI writes that LLMs tend to be overconfident and inconsistent, and puts that overconfidence at the center of Jev’s pitch.

According to TypeSafe AI, the cause lies in training. LLMs are trained with RLHF (reinforcement learning from human feedback), which optimizes them to give answers people like. Jev is trained with a method TypeSafe AI calls RLCD (Reinforcement Learning for Calibrated Decisions): the goal is an answer that comes with a statistically reliable confidence.

I see two concrete uses for Jev. The first is inside applications, to route, classify and rate data without the latency and cost of an LLM. The second is as a middle layer between the interface of a chatbot like ChatGPT and the large language model: Jev’s probabilities and confidence score would catch the answers that sound sure but aren’t.

Can you use Jev yet?

Jev is in early access and isn’t open to everyone yet. You join the waitlist on TypeSafe AI’s website, and the company lets developers in in batches.

If your software makes repetitive decisions, like routing tickets or classifying requests, Jev is a model that also tells you how much to trust the answer. It’s worth joining the waitlist and trying it on your own data.