Jev, System One AI & New Models | Quantized AI 26/08

Jev has attracted plenty of attention with its promise of fast, inexpensive decisions inside software. We take a closer look at what TypeSafe’s System One model does, where it fits alongside LLMs, and what still needs to be tested. We also look at Opus 5.5 and Sol 6.1, and the reported cancellation of Astra 6.1’s planned release over safety concerns. More capable AI is only part of the story; how we use it matters just as much.

The Hype about Jev

Jev is not a new model that is going to replace software engineers completely in the next six months. In fact, it won't even write code. Instead, it is a model for a specific type of task and is meant to be used inside software. That may also be why the company's called TypeSafe AI.

Jev takes a very pragmatic approach. Even before Jev was announced, CEO Diogo Almeida spoke at AI Engineer World's Fair about the limitations he sees in modern LLMs. According to Almeida, training LLMs to please humans does not necessarily make them reliable at automation. In an automated process, the decision needs to be correct, regardless of whether it is pleasing or not. Almeida's argument is that models intended for automation should be trained for that objective.

LLMs are designed to respond to open-ended questions. We can ask an LLM-powered coding agent to write an application that tracks our calories, and it will start to think, develop a plan, execute it and verify the result.

There are also other tasks, though, where the answer comes from a predefined set of options or is simply yes or no.
With common LLMs, users may expect a simple yes or no but receive a lengthy response they have to read through first. And this is where Jev comes in. It is designed to make decisions where the possible answers are known. It doesn’t “think” like the chain of thought in an LLM. You can even pack multiple questions into one request, but it only responds once. Designed specifically for these decision-making tasks, Jev responds faster and at lower cost in TypeSafe’s comparisons. TypeSafe claims speeds 20–200 times faster and costs 40–400 times lower than comparable LLMs. Output tokens are free, so you only pay for the input.

One official example is customer support, where you want to know whether the customer was happy, whether the issue was resolved, and whether the internal rules were followed. That’s multiple yes-or-no questions, with Jev returning a probability for each. Jev can also be used in agentic workflows. For example, it can decide which LLM should handle a user’s input.

There are even skills for coding agents. They teach agents such as Claude Code or Codex how to use Jev’s API. For example, an agent could use Jev to assess whether a particular tool call is safe or needed. That’s a yes-or-no question. If you want an explanation, you could ask an LLM afterwards, because Jev can’t tell you why. That would be the LLM’s own assessment.

TypeSafe calls Jev a System One model, referring to the book Thinking, Fast and Slow. System 1 describes more instinctive, automatic or learned responses - you could think of reflexes - whereas System 2 is where we actively think about something. In TypeSafe’s analogy, reasoning LLMs fall into the System 2 category, while Jev falls into System 1. The name Jev comes from the Jevons paradox, which describes how greater efficiency in using a resource can increase its total consumption.

We already have independent open-weight decision models, such as Laya on Hugging Face. In fact, OpenAI has already announced its Decisions API in limited preview at DevDay.

Our Analysis

Jev is definitely a step in the right direction. It shows that models optimized for specific tasks can be much more efficient.

What does that mean for the future of LLMs? We can also expect the big labs to release their own classification models and integrate them into their agent harnesses. We expect these different model types to be used together, so that from the outside we can’t really tell whether a classification model or an LLM is handling a task. That choice could become the responsibility of an internal router.

Sources

New Models

We also got new model releases. In the last episode, we already mentioned Opus 5.5. According to several of Anthropic’s benchmarks, it seems to lead the models compared. It has much lower standard token prices than its sibling model Fable, which is still on version 5.1. From OpenAI, we have the powerful new Sol 6.1, which, according to OpenAI’s evaluations, comes close to its flagship Astra 6.0 while costing much less. According to The Wall Street Journal, OpenAI cancelled the planned release of Astra 6.1 over safety concerns. Otherwise, we might have had it in October.

Our Analysis

We covered the Hugging Face incident in depth in the last episode. Since then, further incidents involving OpenAI models have been disclosed. We see it as a good sign that OpenAI is applying the brakes and cancelling a planned release over safety concerns.

Sources

Soverius AI

At Soverius AI, we help companies choose, evaluate, and integrate AI models into real software and business workflows. That includes not only selecting the most capable model, but designing the orchestration, safeguards, review processes, and deployment architecture around it.

Quantized AI 26/08: The Hype About Jev

The Hype about Jev

New Models

Soverius AI

Stay Updated

Authors