Jev: The AI Model That Makes Decisions Instead of Writing Text

On September 15, 2026, TypeSafe AI emerged from stealth with a very different kind of model. The company, founded by Diogo Almeida (the OpenAI researcher behind RLHF and InstructGPT), Erik Gafni and Sasha Sheng, raised a $40 million seed round led by DCVC. Its first product is Jev — an AI model that does not generate text at all. Instead, it makes decisions.

Jev is the first of what TypeSafe calls System One models. Where classic LLM APIs work like input → model → generated text, Jev works like input → model → typed decision. You give it some application state and a specific question, and it returns a choice, a score, a probability and a confidence level — something your code can branch on directly.

Decisions, not paragraphs

You ask Jev: Is this transaction suspicious? Which support queue should this ticket go to? Which tool should the agent call next?

Instead of a paragraph your application has to parse, Jev returns a structured answer. If there are five possible categories, it picks one of those five — it cannot invent an arbitrary sixth answer. If it is asked to score a document, it returns a number on a predefined rubric, with per-level probabilities.

The entire API surface is three question types, which is the design rather than a limitation:

• Choice — pick one option from a set you define (up to 255), with the full probability distribution and confidence.
• Score — rate the state along an ordered 2–10 level rubric, returned as a probability-weighted position.
• Yes/no — a simple boolean judgment with confidence.

Multiple questions can be evaluated in a single parallel pass, and every answer ships with a calibrated confidence score. TypeSafe trained Jev with a method it calls RLCD (Reinforcement Learning for Calibrated Decisions), which ensures the model's stated confidence actually matches how often it is right. That is what makes automated workflows safe: high-confidence decisions can run autonomously, while uncertain ones are routed to a human or a more capable model.

Why that is useful

The strongest argument is speed and cost. TypeSafe reports end-to-end latency of 70–500ms and prices Jev at $0.042 per million input tokens (output is free — the model never emits prose). The company claims up to 100–200× faster and 40–445× cheaper than frontier generative models on the small, repetitive decisions Jev targets. It also handles a 64k token context and is available through early access, OpenRouter and the major AI gateways.

That makes a specific architecture attractive, especially for AI agents. An agent constantly makes small decisions: Which tool should I use? Should I retry? Is this action safe? Does this need human approval? Paying for a frontier LLM on every one of those calls is slow and expensive. Jev can sit underneath the agent and handle the cheap, frequent judgments, while a GPT- or Claude-style model is reserved for parts that actually require reasoning or text generation:

Jev for decisions → normal code for deterministic logic → frontier LLM for generation and complex reasoning.

The weak points

Jev is not a replacement for ChatGPT or Claude. It cannot write articles, generate code, hold a conversation or give long explanations. It is built for bounded decisions, and TypeSafe openly documents weaknesses around arithmetic, counting, date comparisons, ambiguous instructions and overly large or noisy context. For calculations and deterministic rules, conventional code is still the right tool.

There is also an important nuance about hallucinations. Jev's constrained output eliminates some forms — it cannot invent an option that was never allowed. But that does not mean it cannot be wrong. It can still pick the wrong answer, and do so with high confidence. The calibration helps you know when to trust it, but it does not make the judgment correct.

The community has also shipped open-source reimplementations of the interface (like openjev and jevlike, built on open models). They are not Jev — the model itself is a closed, managed API with no open weights — but they show the interface is simple enough to clone.

The takeaway

None of this is a new idea on its own. Classification and probabilistic prediction have existed for decades. What is interesting about Jev is that TypeSafe is trying to turn modern AI into a software primitive. Instead of asking a model to write a response and then parsing that response back into data, Jev starts with data and returns a decision.

The future may not be one giant model doing everything. It may be a stack where different models own different jobs — with Jev acting as the fast, cheap decision layer that lets the rest of your application just check a boolean and move on.