Put that way, you could almost think it falls short of ChatGPT, Claude or Gemini.
In fact, it’s trying to do something different.
What interests me about Jev isn’t so much the demo where it plays Mario or the one where it helps pilot a drone. Those examples attract attention, but the real question lies elsewhere: what happens to AI when we stop asking it to talk and use it only to make decisions within an application?
That’s where speed, and especially cost, become interesting.
01
An AI designed to make decisions
Jev was launched on September 15, 2026 by TypeSafe AI, founded by Diogo Almeida, a former OpenAI researcher who worked on the methods that preceded ChatGPT.
TypeSafe describes it as a “System One Model”.
The principle is simple.
A conversational model receives an instruction and generates a response in natural language. Jev works within tighter constraints: you give it context and a set of possible choices, and it returns a decision directly, along with the associated probabilities.
Take a message received by a customer service team.
A general-purpose model might analyse the message and explain that it’s probably an urgent billing issue.
In many applications, that sentence adds nothing.
The software simply needs to know:
- urgent or not;
- billing, technical support or sales;
- automated handling or human intervention.
That’s the kind of decision Jev aims to produce very quickly.
02
When volume changes everything
This is what interests me most.
If you query an AI ten or twenty times a day, a difference of a few cents between two models doesn’t change much.
Within an application, you can quickly reach hundreds of thousands, or even millions, of decisions a day.
Take a simple example: a piece of software needs to make a million small decisions a day and sends around 500 tokens of context each time.
That amounts to 500 million input tokens.
At the price currently listed by TypeSafe — $42 per billion input tokens, with no charge for output — that comes to around $21 a day, or $630 a month.
This is only a rough estimate. A real product will have longer or shorter contexts, different volumes and an entire infrastructure around it.
But the change in scale is easy to understand.
At ten decisions a day, the unit price matters little.
At a million, it can determine whether a feature is economically viable.
03
A product doesn’t always need a chatbot
Take an e-commerce site.
A visitor is looking for “a lightweight jacket for walking in the rain, under €150, not black”.
The site doesn’t necessarily need to reply with a paragraph. Above all, it needs to understand the constraints, decide which products are the best match and in what order to display them.
The same logic applies to a business application.
A case comes in. The system needs to decide whether a document is missing, whether approval is required, whether there’s an anomaly or whether the case can move to the next stage.
Some of these decisions are straightforward to code using conventional rules. Others quickly become difficult to formalise because they depend on context.
An AI like Jev can sit precisely there: between the available data and the action the software needs to take.
I find that approach far more interesting than adding an “Ask AI” button to every tool.
04
For employees, AI could become almost invisible
When we talk about AI at work, we still often think of a new tool to open, a prompt to write or an assistant to talk to.
But in many cases, the end user may have nothing new to learn.
Their inbox could simply do a better job of distinguishing what needs to be dealt with today from what can wait.
A CRM could flag cases that need a follow-up call.
An expense form could detect a missing receipt.
An internal request could be routed automatically to the right team.
The employee isn’t “working” with AI. They’re simply using software that asks them less often to sort, categorise or check repetitive things.
For many employees, the most useful AI may not be the one they talk to, but the one that saves them from making a hundred trivial decisions throughout the day.
05
Mario mainly shows how Jev fits into a system
Demos like Super Mario are useful because they make the process visible.
In the public Mario project, Jev doesn’t play like a human sitting in front of a screen. The program retrieves the game state — position, enemies, environment, movement — and Jev then chooses from several possible actions.
The program observes, Jev decides, the program executes.
In a drone experiment using MuJoCo, we see the same separation: Jev operates at the tactical level, while fast control commands and safety mechanisms remain handled by conventional code.
That’s an important distinction. Jev isn’t an autonomous brain entrusted with the entire system.
Its value is more specific: stepping in where a decision depends too much on context for a simple rule, but is sufficiently constrained for the possible choices to be defined.
06
A structured response can still be wrong
That’s also why I’d remain cautious.
TypeSafe emphasises that Jev produces structured outputs. It won’t suddenly respond with a paragraph when the software is expecting a category.
That’s useful, but it doesn’t mean it’s always right.
It can choose the wrong category or misjudge a situation.
Design work therefore remains central.
What does the system do when its confidence is low? Which decisions can be automated, and which should still require human approval?
An AI like Jev doesn’t remove those trade-offs. Instead, it forces us to make them explicit.
07
I’d test it today, but wouldn’t yet put it at the heart of a client project
Jev is very new.
The figures shared by TypeSafe and the initial usage volumes are impressive, but I’d need more perspective before making it a core component of a business application.
I’d want to see field tests, feedback after several months, documented failures and real use in production.
Today, I’d happily test it in a prototype or a non-critical feature.
I wouldn’t yet build an entire architecture on the assumption that it will deliver on all its promises.
It’s a fairly simple rule with new technologies: a good demo makes you want to try them. It isn’t enough to make a decision.
08
Why its name is well chosen
Jev refers to William Stanley Jevons and the paradox that bears his name.
To simplify greatly: when a resource becomes more efficient and cheaper, we may end up using more of it rather than less.
Applied to AI, the parallel works well.
If a decision becomes fast and inexpensive enough, we won’t just save money on existing uses. We’ll start putting AI in places where it previously made no economic sense.
Searches, forms, rankings, alerts, user journeys, internal tools.
Millions of small decisions rather than a few spectacular conversations.
I don’t yet know whether Jev will become an important technology or simply a very good idea that arrived at the right time. It’s too new to tell.
But if the early results hold up, that’s where I’ll be looking: not at what we can get it to say, but at all the small decisions we can finally entrust to it without sending an application’s costs soaring.
