Why Jev Isn't a Replacement for Frontier LLMs—And Why That's the Point

Who or what the heck is ‘Jev’? While new AI models and tools appear almost every day, only a few cause a real stir - if the promise behind the technology proves true. From a cybersecurity perspective, the TL;DR is that Jev is a “System 1” AI model that can function as a first-layer decision model for fast, high-volume security triage, reducing the number of expensive LLM investigations. At a time of increasing token usage and metered cost, the promise of quick, low-cost triage is very compelling.

System 1 & Human Thought

Originally proposed by psychologists Keith Stanovich and Richard West and adopted by Israeli American psychologist Daniel Kahneman in his book Thinking Fast And Slow, the human brain is described working as two “systems”: System 1 and System 2. While System 2 deals with complex calculations, determinations and problem-solving, System 1 deals with automatic, instantaneous, low effort thought. Kahneman uses the solving of a mathematical equation to illustrate how System 2 thinking works, while showing a picture of an angry human face to illustrate System 1 at work: we recognize the emotion almost instantly, without conscious effort or complex computation.

Jev: A Complementary Probabilistic Layer

Announced in September 2026, Jev’s creators - TypeSafe AI - adopted the System 1 concept to describe their AI model’s operation. Jev is different from conventional LLMs in that it doesn’t generate words for consumption by humans, it outputs decisions that can be acted upon by software. Give Jev a context or situation, and the questions you ask are responded to in a restricted set of ways, including:

• A “yes” or “no” answer

• A precise score

• A choice for a list of predefined options

• A measure of confidence

• An expected probability

For automated systems, especially where AI agents are involved, primitive decision-making is what is required, rather than arbitrary strings of language. This approach could make controlled autonomy more practical because you can put explicit probabilistic decision gates between an AI’s reasoning and its actions. Rather than putting all decisions through a frontier LLM, Jev can take the initial probabilistic guess, routing only uncertain or critical cases for deeper reasoning to the LLM - so, not burning through millions of tokens for low-level stuff.

Looking at current AI security architecture, Jev - and other System 1 models - could function as a complementary, probabilistic layer or wrapper around the traditional LLMs:

• Stage 1: Jev used for large-scale, quick triage, referring less clear, more complex cases to Stage 2.

• Stage 2: Where the frontier LLMs conduct deeper reasoning on the more ambiguous cases, referring the most critical of these to Stage 3.

• Stage 3: A human-in-the-loop makes the final judgement where the stakes, uncertainty or potential impact warrant human oversight.

There is also the issue of cost. At $0.042 per million input tokens and a zero cost for output (supposedly “too cheap to meter” according to TypeSafe), Jev is incredibly compelling. Compare this with standard API input token costs for the likes of ChatGPT, Gemini and Claude, and a similar number of tokens will cost between $1 to $5, depending on tier. While discounts exist for e.g. cached input tokens, the cost of output tokens also needs to be added.

This all sounds great but it’s important to be cautious at this early stage. Jev is not a silver bullet and there are few substantiated 3rd party benchmarks (so far). We're primarily going on TypeSafe's own, plus a bunch of early adopter projects and 'influencer' type videos and case studies. Jev provides probabilities and confidence estimates as part of its output. TypeSafe claims that higher confidence corresponds to higher accuracy and that its decisions are calibrated, but that this is domain dependent. Outside of the domain on which it was trained and evaluated, it will be poorly calibrated - specialized domains in cybersecurity for example.

A Threat To The Frontier Models?

The attractiveness of a faster, cheaper AI model is hard to withstand, and it is arguable that Jev is a threat to the frontier models’ “do it all” approach, including the revenue from token spend this generates. A complementary, probabilistic layer conducting security triage at a massive scale would mean less use is required for the frontier model - or would it? Given the rapid increase in volume and complexity of cyberattacks globally, the frontier models need all the help they can get, and if a System 1 model can quickly triage high-volume, low-risk threats more efficiently, it gives the more traditional LLMs the ability to double-down on the more complex assessments.

It appears that a change in evaluation metrics may be required when it comes to the cost of AI in cybersecurity operations when faced with this expanded architecture. Rather than a cost-per-token measurement, a better method may be to use a cost per resolved security decision. If the adoption of System 1 models becomes part of a shift towards multi-model architectures, a cost-per-resolution metric would be a much more useful way to compare costs, relative to the level of complexity or criticality.

Jev, Meet Laya and friends…

Qwen-based AI models that allow you to run decision logic locally have been available (albeit relatively recently) - including open source tools such as Kev, Nimble, OpenJev and others. Perhaps the speed of AI tool development means that just a few days after the release of Jev, an open-source System 1 model was announced and appeared on Hugging Face called Laya. Some commentators have called Laya a “clone” of Jev, and while we know that web applications can be replicated - albeit sometimes only partially - by frontier model coding tools (Codex et al), it’s difficult to confirm at this stage. While Jev is a closed, hosted model, Laya and others are open-source and licence free. However, like all open-source tools, community development and support is required. [Side note that NVIDIA acquired Hugging Face in early September.]

The Gist

Applying System 1 thinking to create a model that speeds decision-making while reducing compute cost is a practical approach to increasing the efficiency of AI use in cybersecurity. Near term, benchmarks are limited and Jev’s pricing data is early, so we recommend you treat numbers cautiously. Jev is not a one-size-fits-all solution for every domain, but it is an indication that there may be a bifurcation in the way that AI is architected - meaning a separation from the “thinking” part and the “deciding” part. Strategically, Jev should not be positioned as an alternative to the frontier models, rather it is a complementary approach. If we are seeing a shift toward the use of multi-model AI architectures, the winner will be whoever eventually owns the intelligence router - the adaptive and dynamic orchestration system for alerts. For CISOs, the more actionable question is, what is the “fully loaded” cost and effectiveness of these multi-model architectures, and how reliably the system knows when to escalate.