In computer science, **latency** refers to the delay between sending a request and receiving a complete response. On a chatbot, this delay is noticeable as the time before text begins to scroll. While three seconds may seem negligible in casual use, it is significant for software making thousands of decisions per minute. Today, programs often request AI chatbots to respond in formats like **JSON**, a structured checklist. Developers then hope the model does not invent new items and wait for a response. Diogo Almeida spent four years addressing this issue. His solution, **Jev**, is developed by **TypeSafe AI**, a company founded in 2024 in San Francisco. The company raised 40 million dollars in seed funding from DCVC. Jev is described as the first "System One Model," named after psychologist Daniel Kahneman's distinction between System 1 (fast and intuitive) and System 2 (slow and deliberate) thinking. Current chatbots operate in System 2 mode, while Jev aims to function in System 1, reacting quickly without explanation. Jev processes raw situations such as emails, support tickets, or log lines with a list of closed questions, returning answers with probabilities. Unlike classical models that generate answers word by word, Jev calculates all answers simultaneously, enhancing speed. Response times range from 70 to 500 milliseconds, compared to several seconds for large models in reasoning mode. These times were measured from the West Coast of the United States, where the service is hosted. A developer in Nantes would experience additional transatlantic latency. Jev's pricing is significantly lower than other models. It costs 0.042 dollars per million tokens sent, or 42 dollars per billion tokens, with no charge for returned responses. In contrast, large models charge between 0.20 and 10 dollars per million tokens for input and even more for output. These rates are typically seen in smaller, less capable models. TypeSafe AI conducted tests comparing Jev with GPT-6 Astra (OpenAI) and Claude Fable 5.1 (Anthropic), using their average answers as a reference. Jev agreed with the reference in 68% of cases, a level comparable to GPT-5.6 Terra but lower than models like Sol and Opus 5, which achieved around 73 to 74%. Jev does not claim to be smarter but asserts it performs as well as a good intermediate model at a fraction of the cost. Jev's "zero hallucination" claim is nuanced. It cannot answer outside predefined options, choosing among three given choices, never inventing a fourth. However, it may still be incorrect, such as stating "spam at 92%" when the actual answer is different. TypeSafe AI argues that this 92% confidence aligns with 92% correctness, allowing developers to act on decisions above a certain threshold and consult humans below it. Currently, access to Jev is on a waiting list, with only a few users having tested it. Influential accounts promoting the launch noted "paid partnership," suggesting caution. Jev is targeted at developers handling filtering, routing, moderation, or scoring tasks, where manually written rules become too fragile and large models are too expensive. It is not suitable for writing, summarizing, or piloting agents, as it does not generate text. Its choices are capped at 255 options per question, with additional steps slightly slowing performance beyond that limit. There is no public standard benchmark, no published paper, and the company's website is minimal. However, the concept of Jev is considered one of the most interesting ideas in AI development in recent times.