Jev, developed by TypeSafe AI , has launched on OpenRouter. Unlike large text-generating models like GPT and Claude, Jev is defined as the first publicly available System One model: it inputs application states and structured questions, directly returns probabilistic typological decisions, with the goal not to "write answers" but to let the software execute judgments directly.
JEV does not generate text, only makes decisions
Traditional LLM output is essentially a string of tokens. Even when JSON is requested, developers usually still have to handle prompts, parsing, schema validation, and exception outputs. Jev chooses to abandon free text generation, limiting model output to predefined types in the program.
TypeSafe summarizes this as "unstructured state in, typed probabilistic decisions out." Each decision also returns probability and confidence level, and developers can set thresholds: auto execution at high confidence, manual handling at low confidence.
System One targets machine-native AI
The name System One comes from the concept of "System 1" (quick thinking). TypeSafe judges that chat models are mainly optimized for human-AI communication, and that what much software automation truly needs is not beautiful text, but classification, routing, scoring, extraction, and branching decisions.
Therefore, Jev is more like a function call with AI capabilities. It uses TypeSafe's self-developed RLCD (Reinforcement Learning for Calibrated Decisions) training method, aiming to better match probabilities with actual accuracy.
Speed and cost are the biggest selling points
OpenRouter currently offers Jev Latest and Jev 1.13, supports 32K contexts, with an input price of $0.042 per million tokens, and outputs are not billed separately. OpenRouter positions it as a structured decision model suitable for internal software decision nodes such as routing and classification.
TypeSafe's internal workflow tests show that Jev can achieve up to about 193.6x speed improvement and 444.6x cost reduction on System One-type tasks. However, these figures come from vendor-built evaluations, and the team acknowledges that some workflows and testing methods may have biases.
What makes JEV truly interesting is not building a cheaper chat model, but challenging the default premise that "AI models must generate language." If typified decision models can stably enter production environments, the next layer of competition in AI applications may shift from prompt engineering to making models reliable decision-making components within code.
