Jev: TypeSafe AI's System One Model for Fast, Structured Decisions

Introduction

Released by TypeSafe AI in September 2026, Jev is a new type of artificial intelligence model designed to make fast and structured decisions. Unlike traditional large language models that are built to write open-ended text, Jev is engineered to evaluate application states and return typed probabilistic outputs. These outputs include choices, scores, or yes-no probabilities that software can use directly without needing to parse free-form text. The model is named after the 19th-century economist William Stanley Jevons, reflecting the goal that making intelligence cheap and fast will unlock new use cases. The term System One refers to the idea of fast, intuitive thinking as opposed to slow, deliberate reasoning.

Benefits

Jev offers several key advantages over standard large language models for automation tasks. First, it eliminates hallucinations and type errors because it cannot generate text outside its defined schema. This makes it mathematically impossible for the model to make mistakes in data types. Second, it uses parallel processing to score all possible answers in a single pass instead of generating tokens one by one. This results in much faster response times. Third, the model uses a training method called Reinforcement Learning for Calibrated Decisions to ensure that its confidence scores accurately reflect its actual accuracy. This means developers can trust the confidence levels provided by the system.

Use Cases

Jev is designed to act as a fuzzy if-statement for software, fitting easily into existing agent loops and workflows. It is particularly useful for model routing, where it decides which large language model to use based on task complexity. It also serves as an auto mode guardrail to classify dangerous actions before they are executed. A common usage pattern involves a cascade approach where Jev evaluates an input first. If the confidence score is high, the decision is accepted automatically. If the confidence is low, the request is escalated to a more expensive model or a human. It is ideal for tasks like routing support tickets, classifying data, and performing safety checks where the possible outcomes are known in advance.

Pricing

The pricing model for Jev is highly favorable for automation tasks. Input tokens cost $0.042 per million tokens, which is $42 per billion tokens. A major benefit is that output tokens are free. This contrasts with standard large language models where output tokens can be five times more expensive than input tokens. This structure makes Jev a cost-effective choice for high-volume decision-making workflows.

Vibes

Independent testing on the Banking77 dataset showed that Jev achieved 81.1% accuracy compared to 76.4% for a comparable large language model. The model also demonstrated superior calibration with an Expected Calibration Error of 0.082 on choice tasks. Benchmarks on the PriorBench dataset showed Jev completing tasks in roughly 0.35 seconds, while high-effort large language models took about 8.83 seconds. This represents a speedup of 40 times to 200 times depending on the task complexity. Developers have noted that integrating Jev with LangChain allows for easy model routing and guardrail implementation.

Additional Information

Jev represents a shift in artificial intelligence application from generative chat to deterministic decision-making. It is not intended to replace large language models for open-ended generation tasks. Instead, it provides a specialized and efficient layer for the structured decisions that underpin modern software automation. The model supports up to 64k tokens per input, with 32k tokens allocated for state and questions. It supports three primitives for querying: Noul for yes-or-no questions, Choice for selecting from a list of up to 255 options, and Score for rating inputs against an ordered scale. Developers are advised to include an other or unknown option in their schema to handle edge cases since the model cannot generate new categories.