This time, it wasn’t a chatting model that entered the AI agenda, but a decision-making model. On September 15, 2026, TypeSafe AI opened early access to its model named Jev and announced it as the first example of the new model class the company calls System One. The company completed its $40 million investment round led by DCVC.
What exactly does Jev do?
Jev is not a text generator like ChatGPT or Claude. It is given a situation and a set of questions; it returns answers to these questions from among predefined options. Along with the answer, it also gives the probability of each option and the confidence level of the decision. There are three question types: multiple choice, scoring, and yes/no. Multiple independent questions can be asked in a single request, and the model evaluates them simultaneously, not sequentially.
For example, when given a customer message saying they haven’t been able to connect their payment system for 3 days and are losing sales, the model chooses which team should handle it, scores how angry the customer is, and flags whether the message is urgent. All three are answered in a single turn.
Speed and cost claim
TypeSafe announced Jev’s input price as $0.042 per million tokens and counts output as free. According to the company’s own measurements, the model responds in 70 to 500 milliseconds; with text-generating models, the same job can take seconds. The “193 times faster and 444 times cheaper” headline on the company’s homepage comes from TypeSafe’s own measurements. In an independent test, the numbers came out more modest: an events company measured that on 50 real moderation decisions, Jev ran about 5 times faster and 8.6 times cheaper than its competitors, and came out ahead in accuracy with 96 percent.
What tasks is it suitable for?
Jev’s target is not tasks requiring long reasoning, but small decisions repeated millions of times. Routing customer requests to the right team, classifying invoices, scoring security alerts, auditing an agent’s output, or choosing the next move of a browser-using agent are among them. In demos, TypeSafe had Jev play games, choose the path to the target among links on a page, and had it search for flights as the decision engine of a browser agent. The company says that in such tasks, it can reach about 10 queries per second.
Its limits and the point to watch out for
Jev cannot chat, write code, do research, or reason; it also does not accept images, audio, or video. So it doesn’t claim to replace popular chat models. The company’s claim that it doesn’t hallucinate also needs to be read correctly: the model cannot make up a category it wasn’t given and cannot return malformed output. However, this doesn’t mean it will always make the right decision; it can route a message that should go to the technical team to the payment team. So the format is safe, but the decision is not guaranteed.
Jev is currently in early access and is called through a single API endpoint. Recently, on the AI side, answers were being sought for every problem with bigger and longer-thinking models. Jev tries the exact opposite and shows that it can actually be enough for some tasks.
Source: Suat Özkaya · TypeSafe AI documentation

Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.