OpenAI and Amazon Web Services (AWS), among other major players in artificial intelligence, have recently responded to the launch of Jev, a new type of model that sits between large language models (LLMs) and classifiers. The term "clone war" has been used by an alumnus of OpenAI, who now leads a startup called TypeSafe AI. This startup has gained attention since emerging from the shadows in mid-September, largely due to its product, Jev. The name is a nod to the British economist William Stanley Jevons, who proposed that technological efficiency can paradoxically lead to increased resource consumption. Jev is designed to improve the efficiency of using tokens, the basic units of information processed by AI models. TypeSafe AI is promoting a new class of models called System One, which focuses on making "calibrated decisions" rather than generating text. These models provide structured answers to three types of questions: selecting from a list of options, assigning probability scores to a sample, and determining whether a statement is true or false. Each type of question has its own "primitive" (Choice, Score, and Noul), which can be combined in a single query. These are processed separately to prevent confusion between different parts of the input. The name TypeSafe AI refers to the concept of type safety in programming, where a program uses data correctly based on its type. This idea underpins the System One approach, which aims to provide answers that are more straightforward for machines to use than those from traditional LLMs. These answers are more reliable, with built-in confidence scores, faster processing times, and lower costs. For instance, Jev can respond to most queries in under 100 milliseconds at a rate of $0.042 per million input tokens. These models are ideal for tasks requiring quick, human-like responses. They can be used in customer service systems to route requests to the right team, assess levels of frustration, or determine whether a request involves a refund. In more complex systems, they can help with routing, selecting tools, verifying arguments, or applying safety measures. The models can also be combined with deterministic checks in the code for added reliability. AWS has developed its own version of this approach, called Strands Decider, which shares many of the same principles but adds the benefit of requiring less training expertise than traditional classifiers. Strands Decider is open source and designed for local use, based on the Qwen3.5-2B model. AWS has adapted it using a technique called LoRA and replaced the text-generating part with a decision-making component. Performance benchmarks show that the latest version of Strands Decider runs efficiently on various hardware, achieving a median response time of just over 100 milliseconds. OpenAI has also introduced a similar tool called the Decisions API, which uses GPT Luna to handle predefined questions with a limited set of answers. This system can accept images as input, which is an advantage over some other models. However, access to the Decisions API is currently limited, and OpenAI has not developed a specific model for this task, instead relying on existing technology.