For the past few days I've been experimenting with Jev, an AI model from TypeSafe AI that doesn't write text. You ask a question with fixed answer options and get back a choice, a score or a yes or no, with a probability attached. I now use it for my spam filter and for a few sorting jobs I did by hand until last week.
What Jev does differently from ChatGPT
ChatGPT, Claude and Gemini write text. Jev only picks from the options you give it. You send a piece of text along with a question, for example whether a message is spam or how human a paragraph sounds on a scale of 1 to 10, and you have the answer within half a second. You don't get an explanation with it, because that's not what the model is made for.
Vercel put the model on its AI Gateway on September 16, 2026, so you call it with the same key as the other models on that gateway. According to TypeSafe, Jev is up to 194 times faster and 445 times cheaper than a regular language model in their own tests. Those are their numbers. I pay about three US cents for every thousand form submissions it judges.
What I built with it
In one day I put together four things, mainly to see what the model is good at and where it goes wrong.
- A spam filter for my contact form. For each submission, Jev estimates the probability of spam and the probability that it's a real inquiry. Above 90 percent spam, a message quietly goes in the trash. In the gray zone I get the email anyway, with a warning on it.
- A pre-sort of reviews and forum replies. I let Jev run through hundreds of replies on a topic and indicate for each one whether it's usable, what sentiment it expresses and how close the writer is to a purchase. After that I read them myself, but in the right order.
- A score for link opportunities. Jev goes through a list of hundreds of websites and filters out the link farms, and also the big publishers that never respond anyway. What remains are the sites where a mention has a real chance.
- An AI-tell meter for my own copy. Jev gives each paragraph a score from 1 (human) to 10 (machine) and names the writing tics it recognizes. This article went through it too. You can try the meter yourself in the free AI checker, and which tics the meter recognizes is covered in Signs of AI writing: how to spot it and rewrite it.
All in all, the experimenting has cost less than ten US cents so far.
What I'm going to use this for
Until now I used AI to write or to bounce ideas off. Jev is a judge that I can run a thousand times a day without the bill growing. Anywhere I currently make a quick estimate myself before the real work starts, a judge like this can step in: which leads are urgent, for example, or which link opportunities I approach first. I still decide myself, I just start at the top of the pile instead of the bottom.
The model doesn't really read. In my tests it spots a link farm without trouble, but it overlooks a small home decor blog with a page about working together. That's why I only let Jev do the first sift and check the borderline cases myself.
I'll keep playing with it and write down what I run into. You can read how I think about AI in marketing, and why I keep making the decisions myself, in an earlier post.
