Jev is a new AI model developed by TypeSafe AI that is designed to make quick, measured decisions rather than generate conversations or long pieces of text. Launched on September 15, it can help AI agents decide which tool to use, whether to retry a task and how to classify or route information. TypeSafe AI says Jev can return results in less than half a second, with input costing $0.042 per million tokens and no charge for output.

The model was developed by Diogo Almeida, a former OpenAI researcher whose work contributed to instruction-following and reinforcement learning from human feedback (RLHF). TypeSafe AI describes Jev as a “System One” model focused on quick decision-making. It was trained using synthetic data through a method the company calls Reinforcement Learning for Calibrated Decisions.

Unlike large language models such as GPT and Gemini, Jev is not designed to produce open-ended text. Developers set the possible answers in advance, and Jev selects from those options while providing probability scores that indicate how confident it is in each decision.

The model is already being used for tasks such as software automation, classification, model routing and monitoring AI agents. Some early users have reported that Jev can be faster and cheaper than larger AI models for specific tasks. It has also been added to platforms including Vercel, Cloudflare, LangChain and Langfuse. TypeSafe AI has now made Jev available without a waitlist.

What is Jev?

Jev is a decision-focused AI model built by TypeSafe AI. Instead of answering questions with paragraphs, it looks at a set of defined questions and returns structured decisions along with probability estimates.

The model is named after economist William Stanley Jevons, who is associated with the Jevons paradox. The idea suggests that when a resource becomes more efficient or cheaper to use, overall consumption can sometimes increase.

How is Jev different from an LLM?

Most large language models are built to generate text and code. They produce responses step by step, one token after another.

Jev takes a narrower approach. Developers define the possible outcomes, and the model chooses between them. This makes it suitable for tasks such as classifying information, scoring messages, routing requests and making yes-or-no decisions.

Why are developers interested?

Speed and cost are two of the main reasons Jev is attracting attention. TypeSafe AI says the model can make decisions quickly while costing much less than many conventional language models.

Developers are testing Jev for tasks such as checking software commands, sorting business emails, directing AI workloads and monitoring AI agents. Its probability scores can also help developers decide whether an AI system should act on its own or pass a task to another model or a human.

The broader idea is to use Jev as a decision-making layer underneath AI agents. Instead of using an expensive language model for every small decision, developers could use a faster specialised model for simple tasks and reserve larger models for work that requires deeper reasoning or text generation.