What Jev is, and why you should pair it with an LLM

TL;DR

• Jev picks an answer from options you give it. It doesn’t write.
• It’s much faster and cheaper than an LLM, but not smarter.
• Use it next to an LLM, for the small decisions.

---

So, an LLM like ChatGPT or Claude reads what you send it and writes an answer back, one word at a time. The newer ones also write out their thinking before they answer, which is what people mean by reasoning. That’s why they can handle messy stuff, things you can’t decide in one look.

Jev is a different kind of model. It’s made by TypeSafe AI, and they call it a System One model. You send it some text and a question, like with an LLM, but you also give it the possible answers up front. It never writes anything back. You get a choice from your list, a score on a scale, or a yes or no, plus how sure it is. Ask it five questions about the same text and it answers all five at once, usually in about a tenth of a second.

You might ask why not just write a switch case in your code. A switch case needs exact rules, so it breaks the moment someone phrases things differently. Jev reads the text and understands it, it just doesn’t think out loud.

A bit of background. TypeSafe was started in 2024 in San Francisco by Diogo Almeida, Sasha Sheng and Erik Gafni. Almeida spent about four years at OpenAI and co-wrote InstructGPT, the training method behind ChatGPT. Jev came out in early access in September 2026, and it’s named after William Stanley Jevons, the economist who noticed that when something gets cheaper, people use a lot more of it.

Where it’s actually useful

Running a top LLM on every small decision costs a lot and takes a while, and most of those decisions don’t need a written answer. Jev costs $0.042 per million tokens you send it, and the answers are free.

The example you’ll see everywhere is support tickets, and it’s the simplest one to try. A ticket comes in and you want to know if it’s about billing, a bug or a simple question, how urgent it is, which team should get it, and whether the customer sounds annoyed. Jev answers all of that in one call.

A lot of people are also building email classifiers with it and posting them like they found something new. We’ve had email classifiers for years. This one’s cheaper and faster, sure, but how fast do you really need your inbox sorted?

The way I’d think about it: does it need to be instant, and can you answer it without thinking? If you’re going through thousands of LinkedIn profiles to see who went to a certain college, that’s a yes or no. Say 10,000 profiles at roughly 1,500 tokens each: about $0.63 with Jev, and somewhere between $3 and $150 with an LLM, depending on which one. So it is cheaper and faster, but with a cheap LLM you’re paying a few dollars either way. The gap only shows when you do it millions of times, or when it has to be instant. If you want to know whether someone has the soft skills for a role, or you have 100 people and 10 positions to match, you need an LLM that can actually reason about it.

What I tried it on

I tried it on a few things: picking which files an LLM should read, routing tasks to the right model based on how hard or risky they are, judging my X feed while I scroll, and some quick browser tasks. Nothing groundbreaking, but it saved me time and tokens every time. The browser stuff was much faster and the X feed judge worked really well. None of it is going to make you money tomorrow though.

Browser use is the interesting one. When an AI uses a browser, it goes step by step: look at the page, pick the next action (click, type, scroll), do it, check what happened. Picking the next action is exactly what Jev is good at. dax from opencode put it as “browser use is really just a series of classification tasks”. One team took a workflow of about 90 clicks from 18 minutes down to just over 5 by splitting the work. Opus reads the page and plans, Jev picks each click from a short list, and when Jev isn’t sure, Opus takes over again. When something has to be typed, you still need an LLM, because Jev can’t write.

The catch

Since it doesn’t write, it also doesn’t work through context the way an LLM does. It’s bad at counting, maths and dates, it only reads text, and it gets worse when you give it a lot of noise.

People say it can’t hallucinate. That’s technically true, it can’t make anything up if it doesn’t write. But it can still pick the wrong answer and be sure about it, and the founder admits that too.

Same with the speed claims. TypeSafe says up to 193.6x faster than LLMs, but someone went through 12,759 launch tweets and the median person reported about 7x. That’s still a lot.

Fun fact, you can technically make it write. It can’t generate text, but it can pick which key to press next, one letter at a time. It’s slow and bad at it, but it works.

So, does it replace LLMs?

So no, I don’t think it replaces LLMs. LLMs are better at basically everything, but a lot of small jobs don’t need that. They need to be quick and cheap. I’d try letting your LLM call Jev on its own for those small decisions (I think that’s where this is going). It’s new, so I wouldn’t rebuild anything around it yet.

If you want to try it, Jev is on OpenRouter, which is how I got access. And madewithjev.com has a long list of things people have built with it, if you need ideas.

claude · post 01

> explain this like I’m busy

An AI that picks from your options instead of writing. Cheap and fast. Let your LLM hand it the small stuff.

Thanks for reading,