On September 15, 2026, TypeSafe AI, an American startup, unveiled Jev, a new type of artificial intelligence designed to make fast, structured decisions rather than produce text. This model is the first in a new class of systems known as "System One Models," which differ from traditional large language models (LLMs) by focusing on making decisions with probability estimates rather than generating text. According to TypeSafe AI, Jev is not only quick and affordable but also immune to "hallucinations"—a term used to describe when AI systems generate false or misleading information. The model quickly gained interest among developers, with early access made available to all and a waitlist removed. Jev was created by Diogo Almeida, a former researcher at OpenAI and one of the co-creators of ChatGPT. Almeida holds a degree from the Georgia Institute of Technology and previously worked on reinforcement learning from human feedback (RLHF), a method used to train language models to follow instructions and interact with users. He argues that the industry has been focusing too much on conversational models, which he believes are not well-suited for automation. This belief inspired him to found TypeSafe AI and develop Jev. The development of Jev took TypeSafe AI two years, during which the team faced "countless technical challenges and several scientific breakthroughs," according to Almeida. Jev is built to make "rapid and structured decisions" but cannot engage in conversation. Its name is a tribute to William Stanley Jevons, a 19th-century British economist. The startup believes that as the cost of artificial intelligence decreases, Jev will see broader use in various industries. Unlike ChatGPT or Claude, which are powered by LLMs, Jev does not generate text. Instead, it answers three types of closed-ended questions: Choice, Score, and Noul. For example, it could determine if a comment contains offensive language (Noul), classify a message as a question, critique, or spam (Choice), or rate the intensity of the message on a scale from 1 to 5 (Score). These models perform best when each question focuses on a specific, well-defined point. Each query is similar to an intuitive judgment a highly competent person could make in a few seconds, given the right context. To support Jev, TypeSafe AI developed a new technical infrastructure "entirely dedicated to automation," including a "new model architecture" and a "parallel sampler," which allows the model to evaluate all questions in a query simultaneously. The company also introduced a training method called Reinforcement Learning for Calibrated Decisions (RLCD), which teaches Jev to make "calibrated decisions"—essentially, to accurately assess its confidence in the results. Because Jev's answers are constrained by a schema, it cannot hallucinate or make errors. However, it can still make mistakes, and the confidence score helps estimate the risk involved. Jev is priced at $0.042 per million input tokens, with output being free. Internal tests showed response times of 70 to 500 milliseconds, significantly faster than traditional LLMs and much more cost-effective. However, the startup has cautioned that the results might be influenced by the specific tasks used for testing, which were created by its own team.