- Jev is an AI model specialized in returning structured and typed decisions instead of generating conversational text.
- It is based on the 'System One' concept, prioritizing speed and efficiency in sorting, scoring, and routing tasks.
- It drastically reduces costs and latency compared to traditional LLMs by eliminating the generation of exit tokens.
If you spend your days struggling with prompts to get an LLM to return a non-corrupted JSON file or trying to prevent it from freaking out when classifying a ticket , you need to check out what TypeSafe AI is doing. They've launched Jev, a model that completely breaks with the logic of ChatGPT or Claude, because it completely ignores writing sentences and focuses exclusively on making decisions that any program can read without breaking a sweat.
Basically, we're talking about a tool designed for software, not humans. While conventional AI attempts to be a multifaceted assistant that reasons and writes, Jev positions itself as an ultra-fast decision-making layer . It's the missing link between deterministic code (the typical 'if' and 'else' statements) and the semantic power of language models, allowing automation to be much smoother and cheaper.
The philosophy of System One
To understand Jev, one must look to the psychology of Daniel Kahneman. TypeSafe AI has dubbed its creation a System One model . Simply put, System 1 is the fast, intuitive, and almost automatic thinking we use for everyday tasks, while System 2 is slow, deliberate, and costly reasoning. Current LLMs are pure System 2; they try to reason through every word, making them slow and costly for even simple tasks.
Jev does the exact opposite: it specializes in quick, repetitive judgments . Its name is a nod to William Stanley Jevons and his economic paradox, suggesting that when the cost of a decision plummets, the number of decisions we can make explodes. It was developed by Diogo Almeida, who already had significant experience at OpenAI, having worked on InstructGPT, and the project has strong $40 million in funding led by DCVC.
How does it really differ from an LLM?
The difference isn't the size of the model, but how it delivers the information . An LLM generates text token by token, and you have to hope the formatting is correct. Jev, on the other hand, uses parallel sampling. It doesn't write; it chooses. You define the options, and it returns the exact value, eliminating the need to parse text and, by extension, eliminating type errors . If you ask it for a Boolean value, you'll receive a Boolean, not an explanation of why it's true.
Another key point is training. While most AIs are optimized using RLHF to please humans, Jev uses RLCD (Reinforcement Learning for Calibrated Decisions) . The goal here isn't for the response to be "pretty," but rather for the probability associated with the decision to be honest. This means that when Jev says it has 90% confidence, that figure is a reliable indicator of accuracy —something fundamental for deciding whether an action should be executed automatically or if it needs to be reviewed by a human.
The three basic tools: Primitives
All of Jev's magic can be summed up in three types of questions you can ask him in a single request:
- Choice: The model selects one option from a list of up to 255 possibilities. It not only gives you the winner, but also a probability map of all the options and a confidence score. It includes an "other" option to prevent the model from forcing an incorrect answer.
- Score: Assigns a value on a defined scale (from 2 to 10 levels). The interesting thing is that it returns decimal values (like 1.4), which allows for capturing... nuances between categories that an integer would ignore.
- Noul (Evaluation): It's the simplified version of yes or no. It returns a probability between 0 and 1. The closer to 1, the more resounding the "yes".
The most powerful aspect is that you can ask all these questions simultaneously about the same state (text, JSON, or array). For example, in a support workflow, you could simultaneously ask about the problem category, the customer's level of frustration, and whether they are requesting a refund—all in a single network journey and in a fraction of a second.
Performance, costs and the "fine print"
The numbers TypeSafe is reporting are, frankly, insane. They're talking about latencies ranging from 70 to 500 milliseconds and a price of $0,042 per million tokens for entry, with exit completely free. According to their own benchmarks, Jev can be hundreds of times faster and cheaper than front-end models like GPT-4 or Claude.
However, these figures should be taken with a grain of salt. The company itself admits that its tests were conducted internally and that the results may vary. Nevertheless, external tests confirm that, in high-throughput sorting tasks, the savings are enormous. For example, processing thousands of GitHub issues can cost pennies with Jev compared to tens of dollars with generative models, while maintaining similar or even higher accuracy on known paths.
Use cases: Where does it shine and where does it fail?
Jev is the perfect tool for lead routing, content moderation , ticket triage, or creating guardrails for other AIs. A particularly interesting use case is web navigation (as in the jev-ultrafast project), where Jev decides which page element to click without needing to describe the screen, making automation incredibly agile.
However, don't try to use it for everything. Jev can't write code , doesn't generate code, and can't reason through complex steps. Furthermore, it has clear weaknesses: it's not good at math, gets confused by dates, and can fail if the context is too noisy. It's also not a security system; it can say that something "looks dangerous," but the final authorization must be in the code , not the AI.
Technical integration and ecosystem
For developers, getting started with Jev is quite simple. There's an official SDK for JavaScript and Python (version 0.6.0), and it can also be integrated via a simple POST request. Furthermore, it's already available in Vercel's AI Gateway, which greatly simplifies key management and billing if you're already using that environment.
The recommended architecture is a decision waterfall : Jev filters and classifies the massive volume of data cost-effectively, the code handles the deterministic rules, and only the most ambiguous or complex cases are sent to an expensive LLM. This way, you optimize the budget and reduce overall system latency, using Jev as the frontline monitor.
This new category of models transforms AI from a chatbot that tries to mimic us into a surgically precise software tool. By separating decision-making from generational capabilities, Jev enables the construction of much more robust agents, where calibrated trust dictates the workflow and costs cease to be an impediment to large-scale automation.
