Jev: The Model That Ditches Text for Typed Decisions
TypeSafe AI's Jev returns typed decisions 200x faster and 400x cheaper than LLMs for structured tasks.
TypeSafe AI has launched Jev, an AI model that returns precise, typed decisions instead of text, claiming it is up to 200 times faster and 400 times cheaper than a large language model for structured tasks. The company emerged from stealth on September 15, 2026, with $40 million in seed funding.
The core concept is a stark departure from the conversational AI that dominates the market. Jev does not chat. It does not generate stories or answer open-ended questions. Instead, a developer feeds it a shared data state and a set of typed questions. The model returns the correct, typed answers immediately. It is designed for systems that need a quick, calculated judgment, not a human-like response.
TypeSafe AI was co-founded by Diogo Almeida, a former researcher at OpenAI. The startup is positioning Jev as the first commercial "System One Model," a category inspired by psychologist Daniel Kahneman's theory of two thinking systems. System 1 is fast, automatic, and intuitive. System 2 is slow, deliberate, and logical. TypeSafe argues that current LLMs like GPT-4 operate as System 2 engines, which is wasteful for tasks requiring only a quick System 1 decision.
The technology behind Jev is called Reinforcement Learning for Calibrated Decisions (RLCD). Unlike the reinforcement learning from human feedback used to align LLMs, RLCD is not trained on conversation. It is trained to optimize for accuracy and confidence in its direct answers. The model uses a parallel sampler to evaluate multiple potential answers simultaneously, selecting the most probable and confident one. This architecture is built for speed and efficiency in structured environments.
The performance claims are significant. TypeSafe states that Jev is 20 to 200 times faster than LLMs for applicable tasks and 40 to 400 times cheaper to run. The pricing is listed at $0.042 per million tokens, a fraction of the cost of querying a major cloud-based LLM. For developers building high-frequency applications like algorithmic trading, industrial control systems, or real-time data validation, this could represent a fundamental cost and latency reduction.
An independent test by the Italian AI research group OrcaRouter provided mixed but notable results. The test involved diagnosing bugs in a simple program. Jev correctly identified 6 out of 7 defects, showcasing strong capability in that narrow, structured domain. However, the result also highlighted its limits. The model could not explain its reasoning or handle the seventh defect, which required a more abstract understanding of the code's context.
Critics and analysts have been quick to outline the model's boundaries. Jev is not a replacement for an LLM like Claude or GPT. It cannot engage in dialogue, write an email, or reason through an ambiguous problem. Its effectiveness is strictly confined to tasks where the questions and possible answers can be well-defined and typed in advance. It is a specialist tool, not a general-purpose intelligence.
The launch of Jev signals a potential fork in the AI development road. While the industry races toward ever-more capable and general LLMs, TypeSafe is betting on a parallel market for ultra-efficient, specialized decision engines. The real test will be whether developers can build significant products around this new "System One" primitive. Will it remain a niche tool for performance-critical software, or will it spawn a new class of AI-native applications that operate at machine speed without the overhead of language?