Blog›Guides

Jev: The System One AI Model That Decides Instead of Chats

TypeSafe AI launched Jev, a non-chat "System One" model that returns typed decisions with calibrated probabilities instead of prose. Here is what it is, the performance claims, and why it complements generative prompting.

Jev: The "System One" AI Model That Decides Instead of Chats

For three years the AI conversation has been about models that talk: chat assistants that write essays, code, and answers. In September 2026 a startup called TypeSafe AI released something deliberately different. Jev does not chat at all. It does not write prose or code. Given some state and a list of typed questions, it returns typed answers with calibrated probabilities, in a few hundred milliseconds. TypeSafe calls it the first "System One Model," and it is worth understanding why that framing matters for anyone who builds with AI.

System 1 vs System 2

The name is a nod to Daniel Kahneman's Thinking, Fast and Slow, which splits human cognition into System 1 (fast, automatic, intuitive) and System 2 (slow, deliberate, effortful). Most of today's large language models are System 2 machines: they reason step by step, generate long outputs, and are wonderful for open-ended writing and problem-solving. But a huge amount of real production AI is not open-ended at all. It is a fast yes/no or which-of-these decision made millions of times: is this transaction fraud, which queue does this ticket go to, does this image violate policy, what is the next step in this workflow.

Jev targets exactly that System 1 layer. Instead of asking a chat model to "please respond in JSON" and hoping it complies, you hand Jev a typed question with a fixed set of answers and it returns one of them with a probability attached.

How it actually works

The interface is the whole point:

You send state - an invoice, a support ticket, a security alert, a product listing.

You send typed questions - each with a defined set of possible answers.

Jev returns typed answers - the chosen option plus a calibrated confidence score, fast enough to sit inline in a request.

Because the output is constrained and typed rather than free-form text, your code consumes it directly, no prose parsing, no "the model wrapped the JSON in backticks again" failures. And because it is not generating hundreds of tokens of explanation, it is dramatically cheaper and faster on these narrow tasks.

The numbers (and the asterisk)

At launch TypeSafe reported Jev running roughly 193x faster than Claude Sonnet 5 and costing about 444x less than Claude Opus 5 on decision workloads. Those are striking figures, and they come with an honest asterisk: the benchmarks are TypeSafe's own, the workflows were written by TypeSafe's team, and the company says the gains sit at the high end of expected real-world use. Independent verification was still thin in the first weeks. Treat the headline numbers as a vendor claim, not a settled benchmark, and test on your own workload.

TypeSafe is led by Diogo Almeida, a former OpenAI researcher tied to early ChatGPT work, which is part of why the launch drew so much attention.

Why prompt builders should care

If you live in the world of generative prompts, like most of our catalog, Jev is not a competitor to your image and video prompts, or to your ChatGPT and Claude recipes. It is a complement. The pattern that is emerging is a two-speed stack:

System 2 (generative models) for the creative and reasoning work: writing, designing, coding, ideating. This is where good prompting shines.

System 1 (decision models like Jev) for the fast classification and routing around it: filtering, scoring, moderating, deciding what to do next.

The practical takeaway: as you build AI workflows, notice which parts are genuinely open-ended (keep those on a strong chat model and invest in the prompt) and which parts are really just a fast decision (a typed classifier will be cheaper, faster, and more reliable there). Knowing which mode a task belongs to is quietly becoming its own prompting skill.

Read the full breakdown in our wiki: Jev.