Jev model: what can decision AI do, and what does it mean for a company?

What is the Jev model?

Jev is an AI model that chooses from predefined answers to a question and says, for each answer, how likely it is to be right. It was introduced in mid-September 2026 by a startup called TypeSafe AI. According to InfoQ, the company was founded by Diogo Almeida, a former OpenAI researcher. TypeSafe calls Jev a “System One” model: built for fast decisions that software can use directly.

ChatGPT-style large language models (AI that writes text word by word) can phrase any answer. The Jev model can't. In Understanding AI's example, asked “Is this comment spam?”, Jev might give 0.98 probability for yes and 0.02 for no. The article's author replaced Gemini 3 Flash with Jev for spam filtering on their own blog. In the author's view, Jev is faster, cheaper, and makes it easier to handle the cases where the model is unsure.

Why is Jev faster and cheaper than a model that writes text?

The Jev model is faster and cheaper because it doesn't have to generate text: according to TypeSafe AI's announcement, it gives all the answers in a single query, in parallel. The company compares the two with its own figures like this (a token is a small piece of text, roughly a short word or a syllable):

These numbers should be treated with care. On its own workflows, TypeSafe measured results nearly two hundred times faster and more than four hundred times cheaper, but it writes itself that this is probably the upper end of the real gain. InfoQ cites an analysis of user reports after the release: there, the median reported speed-up was sevenfold. TypeSafe also writes that it can't prove the price isn't subsidised, and expects prices to fall over time.

Why is everyone talking about it?

Everyone is talking about the Jev model because it brought a new form, not a new capability. According to Understanding AI, the announcement got nearly 40 million views on Twitter, and more than 100,000 people joined the company's Discord community within a week. According to the publication, Jev showed no new or dangerous capability; it got attention because it is a different kind of model.

The pattern has spread fast since. In early October Cloudflare released open-weight decision models called Clef (they can be downloaded freely and run on your own servers), under the Apache 2.0 licence and, according to the company, compatible with Jev's programming interface. According to TechCrunch, Amazon released its own open-source version at the same time, and at the end of September OpenAI introduced a similar service called the Decisions API as a limited preview. Understanding AI's impression is that Jev performs best for now, but it isn't clear whether the others will catch up soon.

What is the Jev model not good for?

The Jev model doesn't write text, and according to TypeSafe's own documentation it is weak in several respects. On jev 1.13 and its models, the company writes:

• Jev is not a calculator: maths has to be done in code.
• It doesn't count reliably, and it reads dates as text, not as a timeline.
• Its accuracy drops when the input contains a lot of content unrelated to the decision.
• When choosing, it tends to lean towards the option listed first.
• It is most accurate in English; it handles other languages too, but not equally well.

According to TypeSafe, the model can't make a type error, meaning it always returns one of the given answers. So the format will always be right, but the content can still be wrong. That is why a probability needs a threshold, checks and a person alongside it.

What does this mean for a company of 10 to 250 people?

In our view, Jev-style decision models change companies' everyday automation, not chat. In a company's processes, artificial intelligence is most useful in deciding: which category, who to forward it to, whether it can go ahead or needs checking. That step has just become cheap and fast. Because the answer is a number, a simple rule can feed it into the ERP system, the invoicing tool or the CRM.

This is how we would build a decision like this as part of AI automation:

1. 01

   One question, with fixed answers. For example: quote request, complaint, invoice or other?

2. 02

   A threshold. Above 0.9 probability, the system passes the e-mail on; below it, a colleague gets it, together with the model's best guess.

3. 03

   Numbers stay in code. Validation rules check the amount, the date and the VAT, not the model.

4. 04

   A person approves wherever money or a customer is involved. A person presses the button.

5. 05

   A log. Every decision goes into the log with its probability, so later you can see where the model gets things wrong.

We work with Anthropic, Google and OpenAI models on terms without data retention, and run the systems in Hungary, Germany or Finland. We would only put a new kind of model into a company's process, whether Jev or an open-weight alternative, once we had tested it on Hungarian documents and knew where the data goes. That can be checked during the two weeks of an AI Potential Audit: you find out which decisions repeat many times a day, and which one is worth handing to a machine first.