It Doesn’t Talk Like LLMs, but It’s 200 Times Faster: What’s the Hype Around the New Jev Model?

On September 15, 2026, TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, launched a new artificial intelligence model named Jev. Unlike the large language models (LLMs) we are accustomed to, such as ChatGPT, this model does not write text or hold conversations. Instead, it is designed to make fast, structured decisions that software can use directly. The company introduces Jev as the first model in a new category it calls System One Models, aiming to make AI a core part of software infrastructure and smart agents rather than restricting it to a conversational interface. Developing the model took two years and relied on a novel architecture, a parallel sampling mechanism, and a training method the company calls Reinforcement Learning for Calibrated Decisions (RLCD).

The fundamental difference between Jev and LLMs lies in the nature of their outputs and how they operate. LLMs treat language as a sequence of tokens, generating text sequentially, token by token. Jev, however, receives a state or data to be analyzed along with specific questions and then returns structured decisions that code can read and use immediately to execute the next step.

These decisions revolve around choosing from a set of alternatives, providing a rating on a specific scale, and answering binary questions with probabilities, all accompanied by confidence scores. Jev can also process multiple distinct questions within the same request in parallel rather than producing a lengthy text response, eliminating the need to parse text outputs to extract the desired decision.

Thanks to this architecture, Jev focuses on speed and efficiency rather than text generation. TypeSafe states that the model's response time ranges from 70 to 500 milliseconds, making it roughly 40 to 200 times faster than LLMs for tasks framed as structured decisions. In one of the company's own workflow comparisons, it reached speeds 193.6 times faster. The company notes that input costs are $0.042 per million tokens, with outputs provided free of charge, making it 40 to 400 times cheaper than LLMs. This makes it particularly suitable for systems that need to execute a massive volume of small, repetitive decisions.

Jev is not presented as a replacement for general-purpose LLMs, but rather as a component that can work alongside them within complex systems. An LLM can handle the understanding of complex requests, planning, and the generation of text or code, while Jev handles the small, repetitive decisions the system requires during execution.

For example, it can be used to classify an incoming message and determine whether it relates to billing, technical support, or sales; prioritize a customer request; decide whether a task should be routed to a specific agent; or select the tool a smart agent should use next. It can also be used to verify the outputs of another LLM, evaluate whether an API call appears safe before execution, or determine whether a specific outcome requires human intervention rather than allowing the system to act autonomously. TypeSafe presents this concept as a way to build systems that combine an LLM that writes, Jev that decides, and software code that executes.

This approach becomes increasingly important as AI agents continue to evolve. An agent does not need a model capable of writing a full paragraph at every step; more often, it needs rapid, concise answers that determine the next action. If an agent is handling customer requests, an LLM can understand the conversation and draft the reply, while Jev can be used to determine which department the request should be routed to or whether it should be escalated to a human. In another system, Jev could evaluate dozens of conditions and probabilities within a single pipeline, allowing the code to execute the appropriate decision seamlessly.

This combination of speed, cost-effectiveness, and structured output quickly captured the attention of developers following the model's release.

The true significance of Jev lies in the paradigm it proposes: its primary function is to deliver fast, structured decisions that software can rely on directly. If it proves capable of maintaining high accuracy alongside this level of speed and affordability, AI could become a smaller, much more ubiquitous component within applications, operating behind the scenes to make thousands of invisible decisions, while LLMs remain dedicated to tasks requiring deep comprehension, reasoning, and linguistic generation.