18:51Nick Puru | AI Automation · Tutorial
5 Claude Skills To Make Your Websites Look Ridiculously Good
A stacked five-skill Claude Design workflow, and the build-then-critique loop that keeps it from settling for generic output.
August 25thThe argument in one line.
Jev is a classifier-only model that skips text generation entirely, so it lands within a few points of frontier-model accuracy on simple factual questions at a fraction of the cost and time, but falls behind on anything requiring real judgment.
Read if. Skip if.
READ IF YOU ARE…
- You run high-volume routine decisions (triage, tagging, routing) that currently get sent to a full LLM just to pick one of a few options.
- You already use Claude Code, Codex, or another agent harness and want a cheap pre-filter before a request hits a frontier model.
- You're building or operating an AI agency and want a concrete cost and speed benchmark before recommending a classifier model to a client.
SKIP IF…
- You need the AI to write or generate any actual text, code, or content; Jev cannot produce a single word.
- Your decisions require genuine judgment calls rather than a fact sitting plainly in the text; the video's own tests show frontier models still win there.
The full version, fast.
Jev, from TypeSafe AI, is a decision-only model: it never writes text, it just reads a piece of information plus a question and returns one pick from options you supply, along with a confidence percentage. Because it marks every answer at once instead of writing word by word, it runs in under a second and costs about 4.2 cents per million tokens in with free output. Testing it head-to-head against Claude Fable 5.1 and GPT-5.6 Astra across an inbox triage, contract review, AI-slop detection, and lead-scoring task, Jev matched the frontier models on plain factual questions for a few hundred times less money, but its own confidence score drops below 90% exactly where judgment is required, and that's where the bigger models still win.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:47
01 · JEV is going viral
Cold open on the viral launch posts: a Claude Code session token-count drop, a 7-second flight search, and 50 parallel Subway Surfers games, all run through a brand-new model.
00:47 – 01:13
02 · What JEV actually is
Introduces TypeSafe AI, the two-year-stealth company behind Jev, and the three-input structure: information, a question, and a fixed list of allowed answers.
01:13 – 01:50
03 · The three kinds of questions
Jev only answers three question shapes: yes/no (returned as a probability), pick-one from up to 255 options with a confidence score, and a custom 1-10 score that can land between two levels.
01:50 – 02:36
04 · Why it's so fast
Normal LLMs write answers one word at a time; Jev marks every answer on a multiple-choice sheet simultaneously, so a dozen questions on the same text all return together in 0.07 to 0.5 seconds.
02:36 – 03:09
05 · It's just a classifier
Addresses the objection directly: Jev is fundamentally a classifier, but unlike a traditional one it needs no labeled training data. Options are typed in plain English and it works on the first try.
03:09 – 04:37
06 · Getting JEV into your harness
Wiring Jev into Hermes, Claude Code, or Codex takes about a minute: sign up at typesafe.ai or route through OpenRouter, paste in an API key, and tell the agent to install it.
04:37 – 05:17
07 · The numbers: speed & price
TypeSafe's own homepage claims 193.6x faster and 444.6x cheaper. Input runs $0.042 per million tokens with free output; 10,000 decisions at 750 words each costs about 42 cents total.
05:17 – 06:18
08 · Is it actually any good?
TypeSafe's own published test table across four business workflows: Jev is cheapest and fastest on every row and most accurate on none, with the accuracy gap widening as the questions require more judgment.
06:18 – 06:34
09 · How to get access
Two paths in: join the typesafe.ai waitlist (people report getting in within a day) or skip the wait entirely through OpenRouter, which is what powers the live head-to-head comparison tool used for the rest of the video.
06:34 – 09:13
10 · Test 1: the inbox
40 emails styled like a real inbox, three questions each (needs reply, category, urgency) run through Jev, Claude Fable 5.1, and GPT-5.6 Astra. Jev finishes under two seconds for under a cent; the frontier models take 15-30 seconds and stay close on need-reply and category but fall further behind on urgency, the most judgment-heavy question.
09:13 – 10:42
11 · Test 2: contracts
Classifying auto-renewal clauses, notice periods, and risk scores across a batch of contracts: Jev matches the frontier models exactly on the plain-fact questions (does it renew, how many days notice) at roughly five times the speed and a few hundred times less cost.
10:42 – 11:54
12 · Test 3: spotting AI slop
30 homepage paragraphs, half generic buzzword copy and half real writing. All three models catch the AI slop itself perfectly (30/30), but only the frontier models can reliably explain what the specific problem with the copy is.
11:54 – 12:47
13 · Test 4: inbound leads
30 stripped contact-form messages scored for whether they're worth a sales call, which service they want, and urgency. Jev lands one call short of the frontier models (29/30) while finishing in a fraction of the time.
12:47 – 13:14
14 · The pattern across all four
Same result every time: on a simple fact sitting in the text, Jev matches the frontier models for a few hundred times less money. When the question needs real judgment, the big models stay ahead, and Jev's own confidence percentage tells you which rows those are.
13:14 – 13:29
15 · Where to start
Pick one recurring decision (is this lead worth a call, does this review need a reply, which folder does this invoice belong in), write out the options, hand it to Jev, and watch its confidence percentages for a week before letting it act unsupervised.
13:29 – 15:56
16 · Six ways people are using it
Six running use cases: grading competitor ads by awareness stage, a spreadsheet that tags itself by meaning, routing requests to the cheapest model by difficulty, stopping an agent from loading the wrong skill (cut Hermes' wrong-skill rate by more than half), running 50 parallel browsers to find what breaks a release, and finding a specific file by description.
15:56 – 17:01
17 · What to be careful with
TypeSafe's own docs list the failure modes: it can't invent a fourth option so 'not stated' has to be one of your choices, it's bad at math and reads dates as text, it takes your question literally rather than what you meant, and it degrades on long messy documents. A blitz chess match makes the same point visually: Jev won on the clock but lost the game in 18 moves because it decides fast without planning ahead.
17:01 – 17:16
18 · The chess match
The clock-versus-checkmate result closes the loop on what Jev actually is: a fast decision-maker, not a planner.
Lines worth screenshotting.
- Jev cannot write a single word; it only reads information and a question, then picks one answer from the options you give it.
- Because Jev marks every answer at once instead of writing token by token, it answers in 0.07 to 0.5 seconds instead of 3 to 329 seconds for models that write out their response.
- Jev costs $0.042 per million input tokens and charges nothing for its output, since it never generates any text.
- Testing 13 questions in one call came out over 11 times cheaper and almost 10 times faster than asking them one at a time.
- Jev is technically a classifier, but unlike a traditional classifier it needs zero labeled training examples; you just write your options in plain English.
- Across four real business tests, Jev was the cheapest and fastest option on every single row and the most accurate on none of them.
- On a customer-service classification task, Jev scored 76% accuracy against the best frontier model's 78.3%, a 2-point gap for roughly 300 times less cost.
- Jev's own confidence percentage doubles as a trust signal: at 90% or higher it agrees with the frontier models almost every time; below that, a human or bigger model needs to be in the loop.
- On plain factual questions, like whether a contract auto-renews, Jev matched frontier-model accuracy at roughly five times the speed and a few hundred times less cost.
- Jev always picks one of the options it's given, even when the real answer isn't on the list, which is why 'not stated' has to be one of the choices you offer.
- Jev is documented to be bad at math and bad at dates, since it reads a date as text rather than reasoning about which of two deadlines comes first.
- Running Jev as a pre-filter in front of Hermes' 182 skills cut wrong-skill loads by more than half.
- Vercel's CEO reported that swapping their coding agent's per-command safety reviewer to Jev made it up to 18 times faster and more accurate than the model they'd been using.
- In a blitz chess match against a frontier model, Jev won on the clock but lost the game, checkmated in 18 moves, because it makes fast decisions without planning ahead.
- The advertised '444 times cheaper' figure is the ceiling from TypeSafe's own homepage tests, not the average; real savings shrink the moment a task needs even one written sentence.
- A single Jev call can be multiplied across 50 parallel browser sessions testing which UI element breaks a new release, for pennies total.
When a model that can't write still wins
CLASSIFIER ECONOMICS
Jev proves that most of what you send to a frontier model is actually a multiple-choice question in disguise, and pulling those out first can cut cost and latency by orders of magnitude.
02What JEV actually is
- A decision-only model needs exactly three inputs: the source text, a question about it, and the fixed list of answers it's allowed to pick from.
03The three kinds of questions
- Every question a pure classifier can answer reduces to one of three shapes: yes/no as a probability, pick-one from up to 255 options with a confidence score, or a custom score across up to 10 levels.
04Why it's so fast
- Writing an answer word by word is what makes a normal LLM call slow; a model that only marks a box can answer a dozen questions on the same text simultaneously.
- Batching multiple questions into one call ran 11 times cheaper and almost 10 times faster than asking them one at a time in the same test.
05It's just a classifier
- A classifier that requires zero labeled training data changes the economics: type your categories in plain English and it works on the first try, no dataset needed.
06Getting JEV into your harness
- A narrow decision-only model is cheap enough to wire into an existing agent harness as a pre-filter rather than a replacement for the main model.
07The numbers: speed & price
- At roughly four cents per million input tokens with free output, 10,000 decisions of 750 words each costs about 42 cents total, cheap enough to classify everything instead of only what seems worth a person's attention.
- Published multiplier claims like '444x cheaper' come from the vendor's own best-case benchmark; treat homepage multipliers as a ceiling, not an expected result.
08Is it actually any good?
- The real verdict was 'cheapest and fastest on every row, most accurate on none' — a tool can be the right choice without winning on every axis.
- The accuracy gap between a fast classifier and a frontier model widens specifically as a question requires more judgment, which is a useful signal for deciding what to route where.
10Test 1: the inbox
- On fact-based questions the classifier stayed within a few points of frontier-model accuracy in under two seconds for under a cent, but the gap widened on the one genuinely judgment-heavy question in the same test.
11Test 2: contracts
- When the answer to a question is a plain fact sitting directly in the text, like whether a contract auto-renews, a fast classifier matched frontier-model accuracy at a fraction of the cost and time.
12Test 3: spotting AI slop
- Detecting that copy is generic AI filler is a simple yes/no every model nailed; explaining specifically what's wrong with it needs more reasoning, which is where the cheaper model fell behind.
14The pattern across all four
- The same split showed up in every test: matched accuracy on plain facts, a widening gap on anything needing judgment, and the model's own confidence score reliably flags which rows are which.
15Where to start
- Start with one decision you already make repeatedly, write out the exact options, hand it to a classifier model, and watch its confidence percentages for a week before letting it act unsupervised.
16Six ways people are using it
- A pre-filter that scores a request's difficulty before routing it to a model stops you from paying frontier prices for work a cheap model could do.
- Adding a lightweight classifier as a skill-picker in front of a large skill library cut wrong-skill loads by more than half in one real deployment.
17What to be careful with
- A model that always picks one of your given options will still pick something even when the true answer isn't on the list, so 'not stated' or 'none of the above' has to be one of the choices you offer.
- Known weak spots are worth logging before trusting any classifier: poor at arithmetic and treats dates as text rather than reasoning about chronology.
18The chess match
- Speed and planning are different capabilities: winning on the clock isn't the same as winning the game, so it's worth knowing which one your tool is actually optimized for.
Terms worth knowing.
- Jev
- A decision-only AI model from TypeSafe AI that reads information plus a question and returns one pick from a fixed list of options with a confidence percentage, without generating any text.
- System One Model
- TypeSafe AI's category name for Jev, referencing fast, intuitive 'System 1' thinking rather than the slower, deliberate reasoning of a full language model.
- Classifier
- A type of model that sorts input into predefined categories rather than generating free-form text; Jev works this way but skips the usual step of training on labeled examples.
- Confidence score
- The probability Jev reports alongside its pick, showing how sure it is about that answer; used as a threshold for deciding whether to let it act automatically or route to a person.
- Model routing
- Using a small, cheap model to judge how hard an incoming request is before deciding which, more expensive, model should actually handle it.
Things they pointed at.
03:15toolOpenRouter
05:23toolVercel AI Gateway
Lines you could clip.
00:00
“A brand new kind of AI model called Jev has been going completely viral all over the internet this week.”
01:55
“It cannot write a single word. It only makes decisions.”
02:07
“Jev never writes anything. It's just like handing somebody a multiple choice sheet instead of asking for an essay.”
12:13
“Cheapest and fastest on every single row, and the most accurate on none of them.”
16:42
“It just makes very fast decisions. It doesn't plan ahead.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphorstory
So a brand new kind of AI model called Jeev has been going completely viral all over the internet this week. Now this is the launch post right here. There's already more than 35 million views from one of the co -creators of ChatGPT and people they've been testing it like crazy and sharing what they built with it.
Just like this one right here. So this one it cuts a cloud code session from nearly a million tokens to 86 ,000 in just one second. And then this one right here it finds a flight in literally seven seconds for under a cent.
And then this one right here it's playing 50 games of subway surfers at once for literally less than a penny. Now, we have never seen an AI like this.
It cannot write a single word. It only makes decisions. And they claim that it is up to 200 times faster and 400 times cheaper than any other LLMs like CloudFable and GPT -6 Astra.
So what is it and where would you actually use it? Let's get into that. So Jeff, it comes from a company called TypeSafe AI.
They're out of San Francisco, two years in stealth. And this is the first thing that they have actually shipped. And the way that actually works, it's dead simple.
Now you give it three different things. the information first so maybe a contract a customer message maybe it's an advertisement whatever it is and then a question about it like does this renew automatically and the only answers that it is allowed to pick from now it reads from all of that and it hands you back one pick plus a number for how sure it actually is about that there's only three kinds of questions that you can actually ask it a yes or no Does this message ask for a refund?
It comes back as a probability. So instead of a yes, you get 94 % yes. And then from there, you just pick one from a list like which team should handle this, whether that's billing or technical, maybe the sales team and up to 255 options come back and you get a probability for everyone plus a confidence number for the actual pick.
And beyond that, you score it on a scale that you write yourself. How frustrated is this customer from calm to annoyed to fear? And there's actually up to 10 different levels and the answer can land between two of them.
That's the entire product. The reason that this is so fast, it's very simple. Now, normal LLM, like if we take Opus, GPT, any one of them, it writes its answers one word at a time and every single word has to wait for the one before it.
Now, Jev, it never writes anything. It's just like . Now, every answer, it's already going to be on the page, so it just marks the box and it marks every box in the sheet at the same time, which effectively means that...
you can put a dozen questions on the same piece of text and they're all going to come back together. Now, they actually tested about 13 questions in one call and it came out over 11 times cheaper and almost 10 times faster than just going one by one. And that's where the speed comes from and the price because you only pay for what it actually reads.
Now, a lot of people, they have been looking at this and saying like, okay, that's just a classifier. And yeah, it is a classifier fundamentally, but the difference is you do not have to train this. The old way, you would just feed it 10 ,000.
labeled examples first. With Jev, you just type your options in plain English and it works on the first try. Okay, now to actually get Jev in whatever harness that you are using, whether that's Hermes, Cloud Code, or Codex, it takes literally just a minute.
First, you can sign up through typesafe .ai, make your key. You will have to take a little bit to actually get registered because you have to go through kind of a tedious process, or you can just go with OpenRouter, which is what I'm going to be using. Now we simply just have to paste in this command line, or you can just ask AI or Fable or ask...
whatever you're using to help you install this where you just have to provide your open router key and just tell it what you're trying to install which is going to be jeff anyways we will have the command available inside of that guide in our free school community along with a plethora of other things from this video and our other videos as well so Get that there.
All right, now diving into the numbers a little bit further, their homepage, it claims 193 times faster and 440 times cheaper from tests that they ran themselves. Now the price is 4 .2 cents per million tokens going in and every answer coming back is completely free. So let's say you've got 10 ,000 decisions to be making, each one reading about 750 words.
Whether that's a support ticket, a lead form or review, in total, that's just gonna be 42 cents. Now covering the speed, in their test, one decision takes Jeff about four tenths of a second. Tara, this takes only 10 seconds.
Luna, this is... 13 seconds. Sol is 23.
So it's the same job with the same questions. The others, they have to write their answer out. And Rassel, they tested this as well.
They put it in their own coding as a safety check on every command the agent runs. And their CEO said it's up to 18 times faster than the model that they were previously using and more accurate as well. Okay, now you're probably thinking, Nick, like this is cheap and it's fast, but is it actually any good?
So this is the test table that they published. It's about four different business workflows. And the thing that I actually want you to notice is that Jeff is the cheapest and it's the fastest.
this on every row and the most accurate on none of them now on customer service job gets about 76 percent the best model sole gets 78 .3 percent the two -point difference per case jev costs about a hundredth of a cent sold about three cents so you give up two points and you pay roughly 300 times less i would say that's a decent trade -off and accuracy there means how often it actually agrees with astra and fable because that's what they grade against so the way that i would think about this is price per decision not price per token so if a decision is going to be worth less than a cent to you and most of them probably are this is the first model where the ai cost basically rounds to zero okay so how do you actually get your hands on this there's two different ways number one You just sign up on typesafe .ai and you wait for the wait list and people are getting in within a day.
Or you could go to OpenRouter right now. It's the same price. There's no wait list.
I'm just going to using OpenRouter because it lets me put Jeff on the same page as the two biggest models and run the exact same test through all three very easily. So this is the test. It's just one page.
Up here is my OpenRouter key. It stays in the browser. And I ran about four different tests.
Now we ran an inbox one, contracts, website copy, and inbound leads. And there's actually a fifth one where you could just... paste your own text.
And there's only three columns. We have the same rows, the same questions, three different models, Jev, Astra, and Fable 5 .1. All right, so test number one is going to be the inbox test.
So I'm giving it about 40 different emails written to look exactly like what lands in mine. So I gave it sponsorship emails. I gave it some client threads, receipts, newsletters, and some spam, and three questions on every single one.
Does this need a reply from me? What kind of email is this? There's about five different options.
We're a sponsorship, client, receipt, or no. notification. I already covered that.
And how soon does this actually need attention for me? What is the urgency? So I really just care about these urgency emails, anything else, you know, my team can actually handle inside of my inbox, but things like this.
I need to be aware of. And in terms of the urgency, there's about four different stages. No urgency this week, today, or right now.
So let's go ahead and run all three and let's start seeing what we get. So this is already finished under two seconds and this was under a cent. So we have all these different emails.
Let's go to the first one. So no on the receipt or notification. Urgency is this week, this week.
no urgency let's see if there's any urgency for today so it looks like there was one on today this was no it was a newsletter or promotion i need to get back so probably my team member needs to be handling and astra and fable just finished up so fable this took about 30 seconds and 15 seconds for astra let's go down below and let's actually look at some of these some needs reply 33 out of 48 38 out of 40, and 38 out of 40.
So Fable and Astra, they were at the exact same. The category, 38 out of 48, 35 out of 40, 35 out of 40. And then the urgency, 24 out of 40, 35 out of 40, and then 33 out of 40.
Now let's just keep scrolling and see what else we have. Can you trust the numbers? So this is Jeb's own percentage versus how often it was actually right.
We can see here that this was under 60%, 60, 80, 80, 95, 95, and up. And this is exactly what TypeSafe's own table. showed off earlier cheapest and fastest on every single row but most accurate on none and the gap gets bigger the more judgment the question actually needs so you do need to be very wary of what you're handing this off to and what we're seeing with pretty much all of this data is that they're almost all sitting in the low buckets which is what makes the rule work above 90 let it act below that send it to a person or it's to a bigger or somewhat stronger model let's go ahead and run test number two this one is going to be contracts now i've already provided all of the contracts prior, but normally you would just want to provide all of that inside of here.
But of course, you guys aren't going to be using this custom app that I built for this. You would be using it inside of your harness, whether that's Cloud Code or Hermes, or maybe it's going to be Codex. So anyways, let's go ahead and run all three and see what we're getting.
So the three tables that I want to cover. So Jeff. under two seconds again and astra is coming about 15 seconds and three tables that we're going to be classifying is yes or no to the renewing automatically does this agreement renew automatically unless someone cancels it the pick one notice period so how much written notice is required to stop the renewal and then the scoring risk how much could this clause cost the customer if nobody actually reads this so we could see here this is about 18 cents and 27 cents for cloud fable 5 .1 20 seconds about 20 seconds the same for fable so something interesting is renews automatically was 30 30 across the board 30 30 across the board as well for the notice period and the median per row this was under a second this was almost three seconds and you know four and a half seconds for Claude Fable.
So what we're really seeing here is that when the question is a plain fact sitting in the text, like does it renew? How many days? Jev is just as right as the frontier models and it's five times faster and a few hundred times cheaper.
So quite clearly a no -brainer in this hypothetical. Now let's move on to the third one. This is going to be very quick.
It'll just be 30 homepage paragraphs. Half of them are the... buzzword soup that you get when you ask an AI to write your website without telling it anything, leverage, seamless, anything cutting edge, half for real.
Some of it is specific copy. And what we are assigning it here is a yes or no. is it ai slop like does the website copy read like generic ai written filler picking one of the problems what is the biggest problem with this copy and then the scoring so what should the owner do with this copy anyways let's go ahead and run this and see what we get would you have this already finished up under two seconds once again i still can't believe this astra it's coming out to be uh maybe about like 12 seconds or so and fable 5 .1 it is just now finishing up under 21 seconds so similar to the last one about 26 cents and 17 cents for astra and then again under one cent for jeff now if we actually look at all of the data down below right here we can see ai slop uh these were all 30 out of 30.
the problem 14 out of 30 25 out of 30 25 out of 30 so astra and fable were same on that page so it's the same pattern again spotting it jeff but explaining it we have to upgrades to bigger models here let's just cover one more we're going to be doing the leads now test number four we're going to take about 30 messages off of a contact form on an agency website the names and the numbers stripped out and the categories that we are assigning is yes or no is it worth a sales call like is this lead worth a sales call for an ai automation and web agency pick one of the services what is the lead asking for and then the scoring And then the scoring.
So what is the urgency? How soon does the lead want to start? Let's run all three.
Under two seconds again. This one looks like it's going to be maybe like 12 seconds again, but worth the call. 29 out of 30.
Service is 30 out of 30. Urgency is 30 out of 30. So it looks pretty much across the board.
The exact same, more or less. Worth the call. It was only 29 out of 30.
So it was one short of Astra and Fable 5 .1 as well as this just finished up. Once again, you know, sub. 20 seconds now this is going to be the list that a setter calls first it's going to be completely sorted before a person has to read a single one so across all four of the tests it was the same picture every single time when the question is just a simple fact sitting in the text jeff it matches the frontier models at a few hundred times less cost now when the question does need some sort of judgment the big models are still better in jeff's own percentage it tells you which rows those actually are now if you guys want to test this for yourself and get access to all of our other free ai resources It's gonna be completely free, available to download inside of my school community.
Make sure to check that out. But anyways, what I would recommend for you is just to start small. Find one decision that you make over and over again.
Is this lead worth the call? Does this review need a reply? Which folder does this invoice go in?
Write out the options, hand it to Jeff, and watch the percentages for a week before you let it act on its own. So if you're sitting here wondering which decision to actually hand it first, here are six that people already have running that I've seen, and every one of them is the same shape that you just watched it. some text with the question and the options.
And the first one, competitor ads. Say you pull every ad a competitor is running. And this one reads, tired of chasing invoices, get paid in two days, start free.
The question that you give Jeff is what stage of awareness is this ad aimed at? Unaware, problem aware, solution aware, or ready to buy? And it comes back.
problem aware now ask that same question plus the hook and the offer across 700 ads and it's done in under a minute for just a few cents number two a spreadsheet that fills itself in by meaning so let's say that's a column of customer emails you can just add a column called urgency This row says, just checking you got my message last week.
No rush. The scale is no follow -up needed. This week, today, right now, it lands on this week.
And every other row gets rated the same way, about as fast as you can type the heading. The third one is model routing. Now, before a request goes to any AI, Jeff, it reads it first.
Now, this one says, summarize this email in just one line. The question is. how hard is this is it easy is it medium or is it hard easy so it goes to the cheap model the hard ones can go to astro or fable so you stop paying frontier prices for any easy work number four is skill picking for your agent now if you run hermes or cloud code with a big pile of skills it'll likely load the wrong one more often than you would actually expect it to but the turn with this is book a table for four on friday and add it to my calendar now the question is which skill fits this now every one of hermes 182 skills is on this list, plus none of these.
It's answer as you can see here. is it picked the calendar. And TypeSafe ran exactly this on Hermes and wrong loads dropped by more than half.
Number five is testing your website. Now there is dozens of browsers open at once, each one trying to break the new release. Now at every step, Jeb sees the buttons on the page, the login, add to cart, checkout, apply coupon.
And the question is, which one do you click next to break this? You can apply coupon and you can multiply this by 50 browsers and it still costs pennies to run. And in use case number six is finding anything on your computer by describing.
it so maybe you type the pdf that i just downloaded the question is which file do they mean and the options are every file on the machine the newest pdf is at the top of the list in about a tenth of a second and the questions are options for all six of these they're going to be in the same free guide in the community ready to paste into the page as well now the stuff that you need to be very careful with is on their website it actually says zero hallucinations if we scroll down a little bit right here what that means is it cannot answer outside of your list it cannot invent a fourth option but it can still pick the wrong one out of your three and type save says so themselves there's a whole page in their docs on where it actually could go wrong and because it always picks one of your options if the right answer isn't on the list it still picks something so every list it gets none of the above and that's why not stated was in the demo now also from their documentation it's bad at math so keep that in your spreadsheet so it's notoriously bad at dates where it reads a date as text so don't ask it which of two deadlines is going to be coming first it takes your question at face value so
it answers the one that you wrote, not the one that you actually meant. And it also gets worse the more junk that you put in front of it. So a long, messy document is the worst case and it's text only for now.
So there's no images, no audio in about 20 ,000 words at a time. And then the glorious headline of 444 times cheaper. Their blog, it admits that that's the high end of what you'll see in the real world.
Most jobs, it does save a lot less because the moment you need one sentence actually written, a normal model is back in the loop. And I promise just one more, I have to show this off. Somebody put this in a blitz chess match against the frontier model.
One API call per move, Fable was crushing it on the board, but it kept burning six to 15 seconds of move thinking and ran out of time. So Jev, it won on the clock, but Astra, it didn't need the clock. It checkmated Jev in 18 moves with two and a half minutes to spare, which tells you what this model actually is.
It just makes very fast decisions. It doesn't plan ahead. Now to wrap this all up, you have zero reasons to not be using Jev.
I 100 % guarantee that you can be finding somewhere in your operation that you can be using Jev to save you time, to save you money, more importantly, Importantly, but you can just use CloudFable or Astra, whatever you're using to determine where should I be using Jav inside of my operation? It can look at your skills, anything scheduled automations or routines that you have set up.
You can find. all of those nooks and crannies where it can be just more efficient in that manner so 100 go look into that and lastly if you are a business owner you are looking to implement ai inside of your operations but you don't know where to start or what to actually do then there's going to be completely free guide that we are giving away link will be down below in the description on where to actually start and what you should be implementing inside your operations and if you're feeling crazy you can book in a call with our team if you qualify to go ai native so with that being said go check that out if you do not know where to start and we'll show you how to increase your bottom line and drive leverage but anyways thank you guys for watching and i'll see you in the next video
The Hook
The bait, then the rug-pull.
A brand-new model called Jev has gone viral for doing something no other AI does: it never writes a single word, it only picks. Nick Puru tests the hype against three real business workflows and Claude Fable 5.1 and GPT-5.6 Astra to see where a model that can't write anything actually earns its keep.
Named ideas worth stealing.
01:19list
The Three Question Types
- Yes/No — returned as a probability (e.g. 94% yes)
- Pick One — up to 255 options, returns a confidence score for the top pick
- Score — a custom scale you define, up to 10 levels, can land between two
Every Jev call has to be shaped as one of these three question types; there's no free-form output.
Steal forAny place you're currently asking a full LLM to choose from a short list instead of write something.
12:22concept
The 90% Confidence Rule
When Jev's own confidence score comes back 90% or higher, it agrees with frontier-model answers almost every time. Below that threshold, the question likely needs real judgment.
Steal forSetting an automatic-action threshold: let anything scored 90%+ act on its own, route anything lower to a person or a bigger model.
13:21concept
Route by Difficulty
Jev reads an incoming request first and classifies it as easy, medium, or hard before deciding which model actually handles it, so cheap requests never hit frontier pricing.
Steal forA pre-filter in front of any agent or pipeline that currently sends every request to the same model regardless of complexity.
How they asked for the click.
VERBAL ASK
17:10link
“There's going to be a completely free guide that we are giving away, link will be down below in the description... and if you're feeling crazy you can book in a call with our team if you qualify to go AI native.”
Soft verbal CTA at the very end pointing to a free AI Growth Plan landing page (shown on screen) plus an optional qualification call; held until after all the teaching content is delivered.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
