15:11Jack Roberts · Listicle
7 New GitHub Repos You Need Now
Seven fresh GitHub tools for ChatGPT and Claude users, from a chat-controlled flight tracker to a codebase mapper.
September 14thThe argument in one line.
Because Jev only ever has to output a yes/no, a pick from options, or a 1-100 score instead of generated language, it can run narrow business decisions dozens to hundreds of times cheaper and faster than a frontier model, without being able to replace that frontier model's job of reasoning or writing.
Read if. Skip if.
READ IF YOU ARE…
- You already run GPT-6 Astra, Claude, or another frontier model for repetitive classification-style work like filtering, scoring, or routing and are paying frontier prices for decisions that don't need frontier reasoning.
- You build AI agents or internal tools and want a cheaper, faster component model to plug into an existing agentic stack through OpenRouter.
- You manage an inbox, community, or support queue and want a concrete cost comparison before adding an AI classification layer.
SKIP IF…
- You're looking for a general-purpose or creative writing model. Jev has no language output and isn't built to hold a conversation.
- You want an independent benchmark. This comparison is run by the same person promoting his own paid course and tools, using his own test data.
The full version, fast.
Jev is a new model from TypeSafe, a company founded by a former ChatGPT co-founder, built to produce only three kinds of output: yes or no, a pick from preloaded options, or a 1-100 score. It has no chat interface and no general reasoning. Run through OpenRouter against GPT-6 Astra across five tasks, spam filtering, churn-risk scoring, support routing, AI-slop detection, and picking a design system from 300 pages, Jev finished in a fraction of the time and cost, in one case $0.04 versus $15 for the same 1,000 activities. The conclusion: use a model like this when a task reduces to a quantifiable decision, and route anything needing real reasoning to a frontier model instead.
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 – 00:55
01 · Intro
Cold open states the headline claim: a new model that's dramatically cheaper and faster than GPT-6 Astra, with five use cases to follow.
00:55 – 01:53
02 · What Jev Actually Is
Jev is introduced as a product of TypeSafe, founded by a former ChatGPT co-founder, built to answer in exactly three formats: yes/no, a choice, or a 1-100 score.
01:53 – 02:29
03 · Why It Matters
Six stated advantages: fast decisions, low input cost, free output, parallel processing, confidence/uncertainty scoring, and typed answers.
02:29 – 04:29
04 · Level 1
Jev vs Astra on a 1,000-email spam/scam filter, an ownership-identification task, and a buying-signal detector, each showing Jev several times faster and far cheaper.
04:29 – 05:56
05 · Level 2
Extends the same pattern to churn-risk scoring from member messages, support-ticket routing, and intake-form triage.
05:56 – 08:06
06 · Level 3
Jev is used to detect AI-generated 'slop' text via the creator's free Slop Monster tool, judged on words, phrases, punctuation, rhythm, and proof, at $0.042 vs $15 per 1,000 activities against Astra.
08:06 – 09:39
07 · Level 4
Jev picks the best-matching design system out of roughly 300 candidate pages for a client website, at about $1.94 vs $500 per 1,000 runs against Astra.
09:39 – 10:59
08 · Level 5
Jev is used as the routing layer that decides whether an incoming task should go to a light, balanced, or frontier-tier model.
10:59 – 11:48
09 · When To Use Jev
The closing rule: Jev is roughly as capable as a smaller frontier model, so use it only when a task reduces to a yes/no, a choice, or a score, not for reasoning or writing.
11:48 – 11:53
10 · What To Do Next
Points to a follow-up video on increasing capability beyond speed and price.
Lines worth screenshotting.
- Jev only ever outputs one of three things, yes or no, a pick from preloaded options, or a 1-100 score, which is what makes it fast enough to bolt onto other AI agents as a decision layer.
- In a 1,000-email spam and scam filter test, Jev finished in 3.41 seconds for about a cent per thousand emails, versus 9.93 seconds and roughly $5 for GPT-6 Astra.
- Picking the best design system out of 300 candidate pages cost about $1.94 per thousand runs with Jev versus roughly $500 with GPT-6 Astra.
- Detecting AI-written 'slop' cost $0.042 per thousand activities on Jev against $15 on GPT-6 Astra, a roughly 350x price gap for the same classification task.
- Jev has no language output, so it can't write copy, hold a conversation, or replace a frontier model for anything requiring actual reasoning or explanation.
- Jev's raw intelligence is estimated at roughly the level of a smaller frontier model, not a frontier-class reasoner, which is why it's positioned as a component rather than a replacement.
- A cheap, fast decision model can be used to route tasks between a light, balanced, and frontier-tier model, making the routing decision itself nearly free.
- Churn-risk scoring from raw member messages, a task usually run manually or skipped entirely, becomes cheap enough to run on every message when each call costs fractions of a cent.
- The filter for when to reach for a decision model like this: whenever the answer to a task can be reduced to yes or no, a choice among options, or a numeric score.
Not every AI decision needs a frontier-model price tag.
COST OF DECISIONS
A model that only answers yes/no, picks an option, or gives a 1-100 score can run business decisions dozens to hundreds of times cheaper than a general chat model, because it never has to generate language.
02What Jev Actually Is
- A model can be useful without producing language: a model built to return only a yes/no, a choice, or a numeric score can still be fast and cheap enough to matter.
- Reducing a task down to one of three output shapes, boolean, multiple-choice, or score, is itself a design skill worth learning before reaching for a bigger model.
03Why It Matters
- Several advantages can compound rather than trade off: speed, low cost, free output, parallel processing, confidence scoring, and structured answers all showed up together in the same model.
- A model can process many decisions in parallel because each one is independent and doesn't need a shared conversation context.
04Level 1
- A classification task that costs several dollars per thousand runs on a frontier model can drop to fractions of a cent once it's reframed as a scoring decision instead of a chat response.
- The same yes/no or scoring logic applies across very different problems: spam detection, ownership identification, and buying-signal detection all used the identical mechanism.
05Level 2
- Churn-risk scoring, support-ticket routing, and lead-intake triage are all the same underlying task: read a message, then output a category or a score.
- When per-decision cost drops to fractions of a cent, it becomes affordable to score every incoming message instead of sampling or waiting for a human to notice a pattern.
06Level 3
- AI-generated text can be flagged using concrete, checkable signals, words, phrases, punctuation, and rhythm, rather than a vague gut feeling.
- The same slop-detection logic that filters a social feed can be pointed at a company's own marketing copy to check whether it reads as generic before it ships.
07Level 4
- Narrowing hundreds of options down to a best match is a scoring problem, not a conversation, and reframing it that way is what makes a large comparison affordable.
- The cost gap between a decision model and a frontier model widens as the volume of things being compared grows, from single-digit multiples on small tasks to hundreds of times on a large comparison.
08Level 5
- A cheap decision model can be used to manage more expensive models, by deciding which task should go to a light, balanced, or frontier-tier model before any of them run.
- Building a routing layer out of a simple category choice keeps the traffic-control decision itself nearly free.
09When To Use Jev
- The deciding question before reaching for a model like this isn't whether it's smart enough, it's whether the task can be reduced to a yes/no, a choice, or a score.
- A model roughly as capable as a smaller frontier model is still useful specifically because the task doesn't require frontier-level reasoning, only a fast, cheap decision.
Terms worth knowing.
- Jev
- A model from TypeSafe built to output only a yes/no answer, a pick from preloaded options, or a 1-100 score, rather than general text or conversation.
- TypeSafe
- The company that built Jev, founded by someone who previously co-founded ChatGPT.
- OpenRouter
- A service that lets one API key call many different AI models, including Jev, without separate accounts for each provider.
- AI slop
- Text that reads as generic, formulaic, or clearly AI-generated, flagged by patterns in word choice, phrasing, punctuation, and rhythm.
- Model routing
- Automatically sending a task to a cheaper or more powerful model depending on how simple or complex that specific task is.
- Agentic operating system
- A personal software environment that connects multiple AI tools and models together to complete multi-step tasks rather than answer single chat prompts.
Things they pointed at.
Lines you could clip.
00:06
“the world's fastest model and 400 times cheaper than GPT-6 Astra”
01:00
“they've been building in silence with the doors locked and the windows closed for a couple of years”
03:28
“if you were to do this over a thousand emails... you're looking at just under $5”
11:46
“if you can actually reduce anything down to a series of decisions and you can quantify those decisions, you can accomplish some pretty spectacular things”
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
Jeff just dropped, and it is the world's fastest model and 400 times cheaper than GPT -6 Astra. It is a fundamentally different type of AI that you need to understand. And we're going to cover exactly what it is and five incredibly powerful use cases that you can use today to get light years ahead, even if you're a complete beginner, and also when you definitely shouldn't use it.
Now, first thing to understand is that it is significantly cheaper than the frontier models. I'll show you what that looks like in a second throughout this video. Now, six reasons why we actually care.
Number one, it gives you unbelievably fast decisions. Of course, guys, grab the coffee. Very, very, very cheap, super important to bear in mind.
Its outputs are 100 % free. Yeah, exactly, 100 % free. It can do parallel questions, so everything at the same time.
It can actually also measure certainty, as well as give you a very specific output. And also, it looks at typed answers. Now, to really understand...
how Jev actually works. It was created by one of the co -founders of ChatGPT and they've been building in silence with the doors locked and the windows closed for a couple of years. And then Jev popped out and a company called TypeSafe.
It's a fundamentally different way of thinking about AI. Obviously, we only care about what does it actually mean and therefore what changes. Essentially, it can give you an output in three different ways.
It can give you a very simple, yes, do it, no, don't do it. It can give you a choice. So if you preload it with different things that it could choose from, and it can also give you a score.
So a rank from one to 100 and many different ways. So these are the three ways that it can decide what to do. But the cool thing here is that when we combine it with GPT -6 Astra with CloudFable 5 .1 and Jeff, by using some rules and a free skill down below, you can grab completely free.
It'll be the second link. description, we can effectively supercharge, get results way faster that are going to be significantly cheaper. Now with this, Jeff can do some crazy stuff.
It can play games. It can play anything because all it's doing frame by frame is deciding. Off from these different buttons, WASD, which is where to move and also jump, what is the most optimum decision for me to make?
So you can use Jeff to do anything like this, like play games in real time. You can, for example, use it something like this, where it's lightning fast, where it can actually, based on many different classifying criteria, decide what the best image would be or what the best message would be. Now, on top of that, what else can you do?
This is an internal link dealer and look at how fast it's actually filtering it through. But guys, I'm not here just to show you video games. I'm here to show you real -life use cases that are going to blow your mind.
Now, let's kick off with five levels. We're going to go from inbox, community, slop. We're going to do loads of really interesting stuff.
But let's begin with level one. So I pulled together a specific test. And what we're going to do here is we're going to have YouTube comments and we're going to have a series of different tests.
I'm going to compare this with Astra to show you how this works. So this is a spam and scam filter. So effectively what we do is connect Jev to, let's say, it could be anything, right?
It could be, for example, our agentic operating system. Here I can ask questions, power it up and see everything I want to. Or I could just literally connect it to chat GPT or Claude if I want to.
going to do here, I was given it 20 different emails and some are scams and some are spam. And you're going to see how quick they are identifying that. It's going to run both models.
We can see the time here. Look how fast Jev is going. That was complete in 3 .41 seconds.
Astra, it took, well, it's still going to be fair. And I put the scores here so we can identify it, 9 .93 seconds. So it was three times faster.
And if we look at the cost for Jev, that was about one cent. Okay, so it finished three times faster. And for Astra, that was significantly more expensive.
I want to say about it, it's one cent if you were to do this over a thousand emails. And over a thousand emails for Astra, you're looking at just under $5. Let's take a look at this one.
This is ownership. Again, let's run both models and see how it does that. This one is essentially saying, go through my emails and identify who should earn what.
Again, Jev is just significantly faster and it has, once again, a perfect score. It would have cost you two cents with Jev and seven and a half dollars with Astro. You get the idea.
They've got ones for urgency. Got ones, kind of identify. Good buying signal.
And look at how fast that is, guys. 2 .57 seconds, over 1 ,000 emails. That will cost you a cent.
What can you buy with this sense? And again, Astra comes in at 10 seconds and 31. So the point here is that we can actually now just connect and give Astra hands.
So essentially with the skill that I built, we just give it to Astra and effectively when that task could be done by Jeff, we can do it significantly faster and at a way cheaper rate, which takes us nicely onto level two. So could it, for example, in a customer environment, could it identify, I don't know, from a member post exactly how risky they are at potentially churning based on...
what they're saying. So say, for example, you have a CRM with your client and you want to give it loads of data, loads of different data and insights. It could actually calculate a score about how likely it is that an individual is going to leave, that employee is going to hand out a resignation, anything that you want to.
So this here is a churn risk. I've given loads of stuff like, hey, I got my first automation ship. This is a message.
And it's just going to quickly identify how likely it is to do it. This is all fictitious stuff. And again, it was done in 2 .82 seconds and significantly cheaper.
You can do the same thing with member intakes, with support routing, anything that you want to, you can see effectively we're getting the right ticket to the right queue. So we have people dropping messages, asking questions. And let's just say that we need to triage this and say, hey, great.
This is a billing question. This is an access question. This is a tech question.
We can use Jev and Astra together to go ahead and solve that. Same with intake forms, right? It's cool.
I want to build something useful. What's their goal? Are they vague?
Do they build it? Do they want clients? And again, Jev can do that for you straight away.
And again, the big time -safe here, guys, is speed and cost. That's basically it. Speed and cost.
And the reason why it works this way is because Astra... And other models like it, like anything else, like Claude, have been designed to... .
Now, Jev doesn't give you language like that. It's not been optimized in that way. Hence why it can do things in a completely different way.
Now, level three is a completely different one. This is going to be the ability for it to identify AI slop. And by the way, if you're wondering Jev is accessible via open router.
So all we're going to do is just ask for to do that. You give it your open to key and it can call Jev effectively whenever you want to. That practically is how you're going to use it.
Unless you're doing something like this, where you have say an agentic operating system and you're asking questions, I can come down and have a conversation. If you just come down to this free link down below, just give this over to Astra or Claude, and effectively it tells you when you should bring in Jev, when not, and just give it your API key, and then you'll be fully connected.
So let's actually test its ability to identify slop. Now, this is one of my favorite use cases, so check this one out, for example. This guy built a real -time slop detector.
So you know how LinkedIn's been overtaken by slop a little bit, right? Well, this will scroll through Twitter, and you can actually say, hey, great, if you identify anything that sounds like it was written by Claude, Fable 5 .1, just add slop and remove it from my feed you could do this for ads and the reason why this was so revolutionary is because it can do it at a rate that no other model can in terms of its speed and it's so cheap to do that now when you make something very cheap and very fast you genuinely unlock new use cases like this so you could just say hey has this been written by a person if so keep it if not get rid of it at the same time So for example, what we can do here is identify if something is slop or not.
You can do this to identify AI text and everything. So I'm using the slop monster, which is GetRipper that I built. It's free, I'll put a link down below.
It's like built off all the best slop detectors with some additional bells and whistles that make it really cool. So what we're going to do here is let's look at illustrative cost. Now for Astro to do this, you're looking at $15.
Jev is 4 cents, which is crazy. We're looking at words, phrases, punctuation, rhythm, and proof. So let's see if it can identify slop, okay?
I've given it a same text here. Let's run this comparison. real quick and see look at this reads as slop and it correctly identified the punctuation cadence and it came down using my slopology framework okay and it did it at lightning fast speed let's give it another example okay so here's another one for example let's run this comparison see again reads a slop then we come down and test whether or not it's written by a human or was it written by an ai let's go ahead and see if this one works there and again we can try this one more time Now, as you can see, it's likely AI -assisted.
Astra itself can't tell you. So this is another really cool use case. Let's say that you have a website.
I'm just going to throw this in here to my agentic OS for a second, right? This website here, we built with GPT -6 Astra. Now, let's say that you're doing this for a client.
Maybe you want to go ahead and do different things. You can edit it. Whatever you want to do here is fine.
But let's say, for example, you have this design lab, right? And we have a brief. And we wanted to identify, okay, I fed this guy over 300 design systems.
And I'm basically saying, their ability to identify what the best one would be. Well, let's run the comparison and check.
And we can pick any of the Astra reasoning. I'm going with medium just to keep it fair. Again, I'm not even going to extra high.
Bam, it's found one immediately off the different references. It's gone through over 300 pages. That would only cost you $1 .94 if you did a thousand.
Whereas Astra, for example, $500. to do the same thing, which is crazy. Let me give you another brief example here.
If I had a studio portfolio, okay, and I run the comparison, let's see what Jeff does. Jeff found it, and that's what Astra does. Jeff found it at a second.
Ridiculous. Astra is still going. And look, it found the exact same thing, and if you look at the difference...
$1 .95 per thousand and $500. You get the idea. We could use Jerv itself to classify anything you want.
But if this all sounds, guys, like I'm speaking design Spanish, I'm going to put a link down below for you to my full code code and Astra masterclass. You get full access to my entire agentic operating system, as well as my entire courses on learning this technology, building direct system, building beautiful websites, how to use it to get light years ahead of everybody else.
I'll put a link down below before your competitors actually grab it. Say, for example, if it just needs simple and straightforward tasks, it goes to a light model. If it's general tasks with a little more context, we send it to a balanced model like an Opus 5 or 5 .1, something like that.
But then if it needs complex reasoning, we can send it to Frontier. Well, what we can do is add Dr. Jev in the middle of it.
We need to get his doctorate somewhere between the beginning and this part of the video. Now, check this out. What I've done here, for example, is I've given a query.
We're going to run it and we're going to see. basically which models it would rank it and push it through to. As you can see, it's identified.
And if you look, guys, the exact thing, because what I did with this is I basically added a bit of a description. I added all the models, Danhazy, GBT Astro 6, Gemini 3 .8 Flash, Kimi K3, and all of them had a kind of a brief, like, route to this model under these circumstances. And you'll notice that it did that exact thing for us.
in the exact same order perfectly it's it's really crazy let's give another example a 650k video research pack i run it again let's see how fast jeff does again under half a second Astra's going all the way through. Now, the idea here isn't to have Jev solve anything.
What I'm demonstrating with this is that . It's a tool that we're going to use with Astra. And when you give it this prompt, you're able to set that up.
Now, in terms of Jev's intelligence, and this is really important this time, it's about as smart as like a Sonnet 5. So it's like smart. It's quite smart.
It's not frontier level, but we're not using it for that. So the times where we do want to use it is whenever, and this is the key distinction, guys, it is one of these three things. Whenever...
there is a yes now whenever it can pick an option and whenever there's a score now you could use any of the models you want to for this. What we're saying here is it's significantly faster and significantly cheaper.
It's not something that is a chat model. This is a model that's optimized for making incredibly fast micro decisions. And people are blowing it with a creativity on this.
As you can see, people are generating UIs very quickly. Effectively, you're only really limited by imagination. But what I would say to you is very simply this.
If you can actually reduce anything down to a series of decisions and you can quantify those decisions, you can accomplish some pretty spectacular things but it does bring us on to an interesting question and that's that speed and price are one thing but capability is something completely different which is why the next thing we need to do is learn how to 10x your capability by watching this video here together.
The Hook
The bait, then the rug-pull.
The video opens with a bold cost claim, 400 times cheaper than GPT-6 Astra, then spends twelve minutes trying to prove it with five side-by-side benchmark tests instead of just asserting it.
Named ideas worth stealing.
00:59list
Three Ways To Decide
- Yes/No
- Choice
- Score (1-100)
Jev's only three output types: a boolean, a pick from preloaded options, or a numeric score.
Steal forany workflow that needs a fast, cheap classification or triage step
02:29list
Five-Level Jev Test
- Spam & scam filtering
- Churn-risk scoring / support routing
- AI slop detection
- Design system selection
- Model routing
The five business tasks used to benchmark Jev against GPT-6 Astra.
Steal forstructuring a model comparison or an internal AI cost audit
07:19list
Slop Detection Criteria
- Words
- Phrases
- Punctuation
- Rhythm
- Proof
The five signals used to flag AI-generated 'slop' text.
Steal forauditing your own marketing copy for generic AI tells
09:39model
Task Routing Tiers
- Light — simple, straightforward tasks
- Balanced — general tasks with more context
- Frontier — complex reasoning, multi-step, creative work
A three-tier system for sending a task to the cheapest model capable of handling it.
Steal forany agent stack juggling multiple LLMs by cost and capability
How they asked for the click.
VERBAL ASK
09:11product
“I'm going to put a link down below for you to my full code and Astra masterclass”
Soft mid-video pitch for his paid course, framed as an offer for anyone who found the design-system demo confusing
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
