There's a new AI model that just came out. It's called Jeev, and since its release, the whole internet has been buzzing with all kinds of examples showing how fast and powerful it is. In fact, this new model isn't necessarily better than the models we already know, like Astra or Fable.
It’s just that it’s in a completely different category. You can't chat with it or build apps with it, for example; you can only sort information, but 100 times faster and much cheaper. For instance, sorting 500 emails in a few seconds for less than a cent, or navigating a site much faster than with classic LLMs, managing to find us a flight ticket in under 7 seconds.
On paper, it's impressive, but is it actually useful? And more importantly, is it even true? Because often, when there's this much hype online, we wonder if the demos might be faked.
So, that is exactly why I'm making this video. Welcome to Anti-hype, the show where I test the latest trends for you to see whether or not they're worth your time. And today, it's Jeev’s turn.
Let's go. Let's start by understanding the difference between it and a classic LLM like Fable or Astra. When you ask an LLM a question, it responds using a chain of thought.
It takes time because it writes out each piece of information word by word. It’s literally outputting text. Whereas Jeev simply assigns a score to different options.
Typically, if you ask it whether this invoice is real. Where a classic LLM would do it word by word, saying yes, yes, that is the case . Based on what I see, it is indeed a real invoice.
Here, it simply tells us "clean" is the option with the highest probability. Basically, it's an AI that answers like a multiple-choice test, either true or false, or by indicating what it believes is the most likely option. If I ask you what color this suit is, you can tell me it's blue.
Whereas a model like ChatGPT would have looked at the image, done some reasoning, pondered the question, and eventually said, "It seems the person here is wearing a blue suit. " Well, that’s a reasoning model. With a model like Jeev, I would give it, let's say, five options.
Black, blue, red, white, or none at all. And what it will do for me is give me a very high probability for this option, while the probability for the others will be very small. Which will allow me to go very quickly because the response time here is simply 100 times faster than that one.
And so, we might think that it's a model that will only be used to classify text or documents, because it just gives us a probability from a list of possibilities. Except that when you take that reasoning further, you can do things like in a video game, where you can ask: what is the most likely next action to make me win? And so, this allows us to have Jev play video games by imagining the next action, but it also allows us to have it navigate websites, where, for example, they made a comparison between several models.
It had a goal, which was to go from one Wikipedia page to another, and Jev finished it in like 0. 4 seconds, whereas the other models, since they had to read the whole page bit by bit, took much longer. And very, very simply , if we had to understand one thing regarding the difference from the classic models we have today, it's that these models are made for humans.
This means that humans will come and chat with these models, and these models will primarily produce text. Since they produce text, they are optimized to say "Yeah, sure, hi" or write in French or in English. And therefore, give rather long, rather wordy responses because they are optimized for human beings.
Another thing, we know these models can also write code, meaning they can create applications, and so on. So they are optimized for writing code. Except that there is a person, Diego Almeida, a former OpenAI employee who came along —who had, in any case, worked on ChatGPT—who has a different perspective.
He thought, "No, these models are very good for humans and for writing code, but one thing is missing: models for making decisions. " They are models for classifying things quickly. Actually, we realized that most of our models 'usage now isn't just about chatting with models on ChatGPT or similar, nor just creating apps or websites, but really about processing data, long documents, emails, text, classifying, and navigating.
In short, performing actions—automations and workflows that don't require a human-like response, but just knowing the next action or the probability of something. That's where a classification model like Jev comes in, simply because it allows for very quick responses to requests like: "Is this support email an important one I need to reply to? " And most importantly, it does it for much, much less money.
Simply because there's no need for a massive context window or any complex reasoning. Which is what most models like GPT-4 or Claude do today. There's no need to process images or anything like that.
It just needs to respond quickly and simply. How do we go about testing Jev? Well, it's tested via the Jev API.
They’ve set up an API console, so you can sign up and then start using Jev. Depending on when you use it, you'll either be lucky and be able to use it directly in their playground to test, or if their site is saturated, you won't be able to. You'll have to go through other providers like Vercel, which currently offers Jev for free, or OpenRouter, which also provides access to Jev via their API.
But if you have access to TypeChat AI, their playground where you can test things, they have a few examples. Is food a sandwich? Is a hot dog, for example, a sandwich?
Well, there, I get 95%true. Is a car a sandwich? No.
And there you go, that’s how we’ll be able to use it in real life. Except that's a playground, a sandbox for testing. What happened to me is that this afternoon, the TypeChat AI site wasn't working.
So I used Vercel and put in 5 dollars. I've already tested most of the examples found online to see if they were true or not. And I pushed things to the limit a bit to see how far it could actually go.
And I'm going to show you a bit of what I managed to do. The first example is the use case that, for me, is the most interesting: email classification. I use AI, obviously, to reply to my emails, to classify them, and to categorize them.
Here, what I've done is I have a list of emails right here that I can actually click to open. And for each of these emails, I’m going to run Jev on all of them. I'll process them 64 emails at a time because we can process them in batches.
And we’ll see which category it puts each one in, the priority, how it prioritizes them, if it's spam or not, if a reply is needed, and how long it took. So let's launch this and you'll see what the results are. So, I’ll clear the results.
I’m starting. That’s the latency you see, and in the end, it’s still very, very fast. So here we go: AI workshop for your team, work category, normal priority, 3%spam, 96% reply.
So that one definitely requires a response from me. There's no need for a reply there, but honestly, it categorizes them very well and the priority is spot on. For example, high priority on feedback for a workshop here.
High priority, file: a final call before suspension. It classified that one as spam. Okay, it's a fake email anyway.
That one also requires my attention. For this test on 10 emails, it responded in 2 seconds. It answered 10 emails.
Well, it classified 10 emails. Here, I did another example where I take the emails one by one. It processed 24 emails in 55 milliseconds and it cost a total of 0.
002 dollars. Another example: here is a game of checkers where Jeff is white and I am here, and I’m going to start a new game. What will happen is that depending on my move, it receives the whole game in JSON format and tells me what it thinks is the best move.
And based on that best move, it executes it , and it takes about 300 milliseconds each time. Well, I’ll admit I’m more of a chess player. I'm a bit overwhelmed, but you get the idea.
I just wanted a visual example, that's why I did it. I could have made two AIs play together, but I prefer practical use cases. If you're new here, I post videos regularly on AI, automation, and how to improve your digital environment .
So, don't hesitate to subscribe, and if you're enjoying the video, a like is always appreciated. Let’s get back to the video. Another really interesting use case: here, for example, I have a 60-second video where I’m showing my screen.
That right there is the video timeline. As you can see, it's roughly frame by frame. I programmed Jev to go frame by frame and check the probability of confidential content.
Let me explain: I make a lot of videos, and often small email addresses, phone numbers, or other things leak, so we have to be extremely meticulous about what we publish. I just coded something brilliant: I take a video, like this 60 -second clip, break it down frame by frame, perform OCR to extract all the text, and ask Jeff if there’s any confidential content. It didn't find any here, but I ran it on an upcoming video I’m about to release.
You’ll see, it’s quite impressive. There you go. So, this is a real 27-minute video, and as you can see, there's no content here.
But then, hop, it spots potentially confidential content. Here, it sees people's names. It tells me: " Careful, there are email addresses here .
" It says: "There is an email address. " And what I coded also blurs it and highlights the important element right here. It sees an email, and what’s great is I have categories like keys, IBANs, cards, or whatever.
the device, one email, three emails here that it managed to detect. It's just magical. And honestly, I think this is going to save me a lot of time because I won't have to stress about blurring and can just run a JV pass at the end and it'll be sorted.
I'm so, so happy to have been able to do this because it reads it frame by frame and shows me exactly where the confidential data is, and I think it's brilliant. And what's crazy is that it analyzed this whole thing. It's a 27-minute video.
It analyzed it in a minute and a half. A minute and a half. Whereas my team and I spend so much time watching frame by frame to make sure we haven't missed an API key or anything that's leaking, or an IBAN being exposed, and this is going to save us a lot of time.
Another example is pre-qualifying inbound leads. So, basically, all the business proposals we receive. We get a lot of spam nowadays, for instance.
So what I did is I put in 100 examples containing proposals. And what I'm going to do is analyze all these proposals to see which ones are bad inbound. So, based on certain criteria I defined beforehand, it will be categorized accordingly.
For example, a good inbound. Here, I am Mayac, in charge of partnerships. There's a specific deliverable, a stated budget, explicit sponsorship, a defined budget, a schedule, and clear payment terms.
That makes it a good inbound. And here in my whole list, these are just examples, of course, it gives me a confidence score and a duration. So basically, per email , it takes 200 or 300 milliseconds, which is just insane.
And where a classic LLM—when I use Herm—the way it processes emails is by sending them to a model like GPT or Fable. They cost much more, but more importantly, they are much slower. So I could never process my entire inbox.
What I saw on the internet was like 1,700 emails processed for about 19 cents in a few seconds. And that is just unbeatable. So honestly, for this use case of processing emails or replies, it's just wild.
So it works with emails, works with a lot of things. Here I did it for text messages, where, is this text a scam? I created three categories: suspicious, scam, legitimate.
I run it here, you see for 541 tokens it's only 0. 002 dollars and it handles it. It manages to categorize them very well.
So, I also did other examples, for instance customer reviews. In the same way, it categorizes them even faster, taking each review one by one to see if it's mixed, happy, or furious. Same for job offers, we can categorize them into three options, really depending on the options you provide.
I encourage you to go look at this walkthrough. It explains a bit that there are several ways to perform these evaluations. Either it's based on a choice, a score, or how much it is a yes or no.
I suggest checking the documentation if you're interested, or simply giving it to your agent and it will code these things. It really didn't take me long at all to create these examples. The one that took the most time was the video editing one.
But other than that, it was very fast. Now, let's talk about the limitations. Where AI messes up, because not everything is perfect, there are times with traps it can't grasp.
If you go to the Jeve doc, it tells you a bit about when things bug out. Here, what I did is I put the trap here in the question. So, typically, how many times does the letter N appear in "anticonstitutionnellement" five times?
So, it's supposed to tell me yes , but you'll see what it actually tells me. So here, basically, according to it , it's a yes at 59%. Clearly, 59%isn't 90%.
It's not entirely sure. Likewise, does this sentence contain exactly how many words? So 1 2 3 4 5 6 7 8, there are 18.
Well here, it thinks it's a yes at 53%, which is clearly not enough to give a clear answer. And there, it missed the mark. For example, is January 10th, 2026, a Saturday?
Actually, it's a Thursday, and it got it wrong. So you'll have to pay attention to the different probabilities. When we're at a probability below 75 or 70 in my opinion, well, maybe double-check with an AI, another AI to ensure we're making the right choice.
There’s another thing I did, which is actually comparing it to different models. If I compare them one by one, you’ll see the difference in speed. So here, one by one, it goes through Jev and on the other side, through GPT 5.
4 mini. Jev is obviously much faster. Here, you see , it goes one by one because, well, GPT 5.
4 Mini has to respond, even if it responds with a text message, it takes much more time. There are times when it messes up. So here for example, it messed up on this email.
But then, I have to pick a model that's more robust , like GPT 5. 5. There’s no contest, meaning this one goes much faster.
GPT 5. 5 per email takes like 6 seconds, 7 seconds, there you go, 7 seconds per email. So it's not at all the same thing because it does a whole reasoning process, a chain of thought, which we obviously don't need with Jev.
Those are the few examples, friends, that I wanted to show you. Actually, in cases where you run workflows like me for email management where you need automation workflows, which I believe is the case for many people, I think we'll be able to replace them with Jev. But only when we have either a speed or a budget issue, because for many of us, we have a Claude or ChatGPT subscription which is generally enough for most of our tasks.
But often when we process data in batches, meaning a large quantity, we quickly run out of tokens or it eats into our classic subscription, and that’s where Jev becomes very interesting to use. What you can do is ask your Claude, your Open model, or your ChatGPT: when can we switch models and use Jev? And you will see that based on your usage history, it will find use cases that will be very interesting for you.
So there is a prompt to type that I’ll put in the description, which is very, very simple and will give you ideas on whether or not you should use Jev. For me, it’s handling emails, support tickets, and YouTube comments to see if there are any ideas for new videos or not. Using it for my YouTube edits to check if data is leaking, or just browsing web pages and doing some scraping.
It’s just crazy what this will allow us to do, and we’re only at the beginning. I just discovered this thing this afternoon. So I’m going to spend all my time geeking out on it even more.
I’ll keep you posted . If you want me to show more examples, because this was just roughly an afternoon of work. If you want me to spend more time testing it on my real-life use cases and show you, don’t hesitate to tell me in the comments; it would be my pleasure.
So, that was episode two of Anti-Hype. I hope you enjoyed it. If you haven’t seen the first one, don’t hesitate to go check it out.
As for me, I’ll see you in another video. Ciao. Ciao.
