Video walkthroughs · Claude Code Club

What Is Jev AI? (And Is It Better Than Claude?)

12 minute read · Companion guide to “Jev Sucks, Actually”

Jev AI is the new model everyone is posting about. The short answer to the question most people are typing: no, Jev is not better than Claude, because it does a different job. Claude writes, codes, and talks with you. Jev picks one answer from a list, very fast and very cheap. This guide explains what Jev is, what it costs, what it is actually good for, and how to decide if it belongs in your workflow at all.

Watch the full breakdown: every big Jev demo, tested on one question. Would a normal person actually use it?

What is Jev AI?

Jev is the first public model from TypeSafe AI, a company founded by Diogo Almeida. Before TypeSafe, Almeida was a researcher at OpenAI and worked on the methods that made language models good at following instructions, which became part of the research behind ChatGPT. TypeSafe launched Jev on September 15, 2026 and calls it a "System One Model."

That name borrows from a simple idea about how people think. System one is the fast, gut-level kind of thinking. You glance at an email and you already know it is spam. System two is the slow, careful kind, like writing a proposal. Jev is built only for the fast, gut-level kind of decision. It looks at an input and makes a choice. It does not reason out loud, and it does not write anything.

In plain words, Jev is a classifier. A classifier is a tool that sorts things into buckets. You give it something to look at, plus a question and a fixed list of possible answers. It hands back a score for each answer, showing how sure it is about each one.

Is Jev better than Claude? The direct answer

For almost everyone reading this, Claude is the better tool, because Claude can do the actual work. Jev can only make a choice about the work. The two models are built for different jobs, so the honest comparison looks like this:

• Writing and coding: Claude writes posts, emails, pages, and code. Jev writes nothing at all.
• Explaining itself: Claude can tell you why it chose something. Jev returns scores with no explanation.
• Speed: TypeSafe says Jev answers in 70 to 500 milliseconds. Frontier chat models can take anywhere from a few seconds to several minutes on the same kind of task.
• Cost: Jev charges $0.042 per million input tokens and nothing for output. That is far cheaper than any Claude model, including Haiku.
• Pictures, video, and audio: Claude can read images. Jev works on text only.
• Math, counting, and dates: Jev can't do them. Claude can.

So the real question is whether you have a job where you need thousands of quick, cheap choices from a fixed list. If you don't, Jev has nothing to offer you yet. If you do, Jev can sit next to Claude and handle that one narrow part.

What Jev AI can and can't do

Here is the full picture, pulled from TypeSafe's launch post and from testing the public demos. Jev is good at three things:

• Labeling: is this email a lead, a sponsor, or spam?
• Routing: should this task go to the cheap model or the smart one?
• Rating: how likely is this reply to be a complaint, on a clear scale you set?

And here is what it can't do. It can't reply to you in words. It can't write code. It can't tell you why it chose an answer. It can't see pictures, video, or sound. It can't do math, count, or work with dates. TypeSafe describes this as "giving up strings," meaning it gave up generating text so it could be fast, cheap, and always return an answer in the exact shape you asked for.

That last part is TypeSafe's biggest claim: because Jev can only pick from the answers you list, it can't make up an answer that isn't there. That is a real advantage when software needs to read the result. It is also the source of Jev's biggest weakness, which the next section covers.

How a Jev schema works, and where it breaks

Every Jev request needs a schema. A schema is just the question plus the list of allowed answers. The example from the Greg Eisenberg podcast with Ryan Vogel makes it clear. The input is a description of an iPhone. The question is "what color is this phone?" The allowed answers are blue, orange, red, green, and yellow.

Jev reads the input and returns something like 80% orange, 10% red, 10% blue. It never says "orange" as a sentence. It gives you odds across the answers you allowed, and your software picks what to do with them. TypeSafe says Jev supports up to 255 answer options in one list.

Now picture that the phone is black. Black is not on the list. Jev can't answer "black" because you never gave it that option. It will still hand back a confident-looking score for one of the five colors, and that score will be wrong.

A simple guard helps. Always add a catch-all answer such as "other" or "not sure," and send anything that lands there to Claude or to a person. That one extra option turns silent mistakes into items you can review.

Is Jev AI free? What it actually costs

Jev is not free, but it is close. TypeSafe's official price is $0.042 per million input tokens, or $42 per billion. Output tokens cost nothing, which TypeSafe calls "too cheap to meter." A token is a small chunk of text, roughly three quarters of a word, so a million tokens is a lot of reading.

For comparison, TypeSafe lists frontier model input prices at $0.20 to $10 per million tokens, with output around five times more expensive than input. In the video's comparison, that works out to Jev being about 24 times cheaper than Claude Haiku and far cheaper than larger models like Fable.

You sign up at console.typesafe.ai and the documentation lives at docs.typesafe.ai. New accounts have been reported to get a small amount of free credit to start, so check the console for the current offer before you plan around it.

Is Jev legit? How to read the speed and cost claims

Jev is a real, working model from a real company with a well-known founder, and the demos are real. What deserves care is the headline numbers. The "20 to 200 times faster" and "40 to 400 times cheaper" figures come from TypeSafe's own tests. As of launch, we could not find an independent benchmark that checks them.

TypeSafe is open about how it tested. In its launch demo, the chat models had reasoning turned off to keep the demo watchable. In its accuracy tests, the "right" answer was the answer GPT-6 Astra and Fable 5.1 gave, so the tests measure how often Jev agrees with those models. Both details matter. Speed and accuracy numbers depend on exactly what you compare against. The speed gap will look smaller against a small, fast model and larger against a big reasoning model.

What is Jev AI used for? The use cases people have tested

In the first week, builders posted a flood of Jev demos. Here is how the popular ones hold up for someone running a real business, from least to most useful.

• Sorting 150,000 Skittles by color. A great speed demo. It shows how fast Jev is, and it has no business use.
• Cleaning out an inbox. Jev labels each email as a sponsorship, a lead, or spam almost instantly. Handy, but most people only need to clean an inbox once.
• Tagging YouTube comments. Sorting comments into praise, criticism, questions, and video ideas. Useful only if you then act on the tags, like answering every question.
• Playing fast video games. Jev making split-second moves in a Smash Bros style game. Impressive proof of speed.
• Running a self-driving car simulation. Jev making every choice at stop signs, lanes, and crossings. This is where Jev shines: thousands of quick decisions every second.
• Routing tasks to the right model. Jev reads each task and sends simple ones to a cheap model and hard ones to a smart model. For a normal builder, this is the most useful Jev idea we've seen.

Notice the pattern. Labeling on its own does almost nothing for you. The value always comes from what you do next with the label. If a demo stops at the label, it is a speed test, not a workflow.

Can you use Jev with Claude Code?

Yes, but not as a replacement for Claude. The useful setup is a model router. A router is a small step that runs before each task and decides which model should handle it. Jev reads the task, picks "simple" or "complex," and the task goes to Haiku or to a bigger model. The same idea works for picking which skill should run, using each skill's short description as the list of answers.

TypeSafe does not ship an official router. The Jev model routers you'll see online are built by community developers on top of the Jev API. Some plug into Claude Code directly. Read the code, or ask Claude to read it for you, before you install one, and make sure it falls back to your normal model if Jev fails.

Also know that you don't need Jev to route models. You can tell Claude Code which model to use for which kind of work today. Jev only makes the choosing step faster and cheaper, which matters most when you run a very large number of tasks.

The rule for deciding if any new AI model is worth your time

A new model drops almost every week, and each one shows up with a headline number. Here is the test we use, and it works for Jev and whatever comes next. Ask what job the model does, not how fast it does it. Then ask whether you have that job, today, at a volume where the gain would matter.

Jev fails that test for most creators and small business owners right now. You don't have millions of sorting decisions to make every day. You have posts to write, offers to build, and customers to answer, and that is Claude's job. Jev passes the test for people building agents, games, simulations, and big data pipelines, where a machine makes huge numbers of fast choices.

The same test protects you from the churn. Most of the time, the right move after a launch is to keep your workflow the same, watch what builders do with the new model for a few weeks, and adopt it only when someone shows a use that matches your real work.

How to try Jev in one sitting

If you have a sorting job and want to know if Jev fits, this is the shortest honest path to an answer.

1. Pick one real sorting job you do by hand today, like labeling incoming emails or support messages.
2. Write the question and every possible answer. Add "other" as the last option.
3. Collect 50 real examples from your own inbox or files.
4. Sign up at console.typesafe.ai and follow the quick start at docs.typesafe.ai, or ask Claude Code to write a small script that sends your 50 examples to Jev.
5. Check every answer by hand and count how many are right.
6. If fewer than 45 of 50 are right, rewrite your list of answers and run it again.
7. If Jev gets them right, send the "other" answers to Claude and let Jev handle the rest.

You're done when you know your real accuracy on your own data. That number, not the launch chart, tells you whether Jev earns a place next to Claude.

Read this online at claudecodeclub.ai