How TypeSafe's System 1 model Jev cuts LLM costs by 70% by handling routing, skill picking, and high-volume classification at 4 cents per million input tokens.
You build autonomous agentic workflows or code with Claude Code and are burning expensive Opus or Sonnet tokens on simple decisions
You operate high-volume business automations (email classification, spam detection, invoice verification) bottlenecked by LLM latency and cost
You manage large prompt workspaces with dozens of agent skills and want faster, deterministic skill retrieval
SKIP IF…
You only use AI models to write prose, long-form copy, or open-ended generative text
Your workflows process low volume where API token costs are negligible
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.
Demonstrating Jev's sub-second classification speed and introducing the premise of integrating it into agentic harnesses.
00:32 – 03:00
02 · What is Jev? Architecture & Pricing
Explaining Diogo Almeida's background, the $0.042/1M input token pricing, free output tokens, and the 3 allowed answer shapes.
03:00 – 05:00
03 · System 1 vs. System 2 Mental Model
Contrasting fast-classification System 1 models with token-by-token generative System 2 LLMs using Kahneman's cognitive framework.
05:00 – 07:23
04 · Setup & API Configuration
Accessing Jev via TypeSafe console or OpenRouter and connecting it to Claude Code with starter prompts and masterclass resources.
07:23 – 08:01
05 · Community Interlude: RoboNuggets Pitch
Highlighting student case studies and community courses on AI agent development and monetization.
08:01 – 10:44
06 · Level 1: Model Routing & Skill Selection
Implementing automated model routing to save 70% on token bills and indexing 145 workspace skills in 5 seconds instead of 30.
10:44 – 11:46
07 · Level 2: High-Volume Business Automations
Benchmarking 100 emails for lead temperature triage and detailing enterprise use cases like fraud detection and churn prevention.
Atomic Insights
Lines worth screenshotting.
Jev does not generate text word-by-word; it outputs structured classifications in a single pass in under 0.3 seconds.
TypeSafe charges zero dollars for Jev output tokens, pricing only prompt input tokens at approximately $0.042 per million.
Jev constrains answers to exactly three shapes: binary boolean, multiple-choice selection, or numeric scale rating.
Categorizing AI into Daniel Kahneman's cognitive framework: Jev represents intuitive System 1 snap judgment, while autoregressive LLMs (Claude, GPT) represent deliberate System 2 thinking.
Automating model routing with Jev cut token expenditure by 70% across 12 test tasks because 9 out of 12 prompts never needed a flagship model.
Querying 145 workspace skills directly through Opus 5 took 30 seconds, whereas routing through Jev retrieved the correct skill in 5 seconds.
In business triage benchmarks across 100 customer emails, Jev classified lead temperature in under one second, beating Haiku by 14x and Fable by 16x in speed.
High-volume classification tasks such as invoice fraud, refund validation, and customer churn scoring become economically viable at scale when output generation cost is zero.
Takeaway
Hybrid agentic architectures beat monolithic LLM prompts on cost and latency
CORE ARCHITECTURE
Autonomous agents waste massive compute generating open-ended reasoning tokens when all that is needed is a categorical routing decision.
02What is Jev? Architecture & Pricing
Constrain decision-making tasks to non-generative architectures to capitalize on zero-cost output tokens.
03System 1 vs. System 2 Mental Model
Separate fast intuition (System 1) from deliberate prose generation (System 2) rather than forcing one model to do both.
06Level 1: Model Routing & Skill Selection
Automate dynamic model dispatching; over 75% of routine agent prompts can be resolved by low-cost models like Haiku.
07Level 2: High-Volume Business Automations
Run bulk classification pipelines through single-pass models to eliminate latency bottlenecks in high-volume queues.
08
Deploy lightweight classifiers client-side or at the edge to enable real-time UI filtering and semantic search.
Glossary
Terms worth knowing.
Jev
A proprietary frontier AI classification model released by TypeSafe AI, designed specifically for rapid decision-making rather than generative text output.
System 1 vs System 2 AI
An architectural framing based on Daniel Kahneman's cognitive psychology: System 1 refers to sub-second snap classification models, while System 2 refers to token-by-token reasoning LLMs.
Model Routing
The programmatic practice of evaluating an incoming prompt to determine the smallest and cheapest viable model required before dispatching the execution call.
Agentic Harness
The surrounding software environment (e.g., Claude Code, OpenClaw, custom agentic OS) that equips an LLM with tools, file access, and execution scripts.
Free Output Tokens
A pricing model where the API provider charges exclusively for prompt ingestion while classification responses incur zero token cost.
“its output tokens, basically its response to you as the user, is always free. And they only charge for the input tokens”
Summarizes the disruptive economic model behind Jev in plain terms.→ newsletter pull-quote↗ Tweet quote
01:53
“Jev is the first System 1 model, whereas all the other AI models like Fable, Astra, and the others are what they're calling System 2 models.”
Clear conceptual distinction framing the evolution of AI architectures.→ TikTok hook↗ Tweet quote
06:04
“it actually resulted to 70% savings because nine out of those 12 tasks never needed the top model anyway.”
Concrete ROI metric proving the financial benefit of automated model routing.→ IG reel cold open↗ Tweet quote
The Script
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.
17px
metaphoranalogy
There's a new AI model in town from the co -inventor of ChatGPT and it is insanely cheap and incredibly fast. It's called Jev and just to show you how fast it is, I'll send this prompt. And that is real time.
It gave me an output for less than a second and at a fraction of the cost. So today I'll explain Jev for you simply and also some of the best ways by which you can integrate Jev with the agentic harnesses you use, like Cloud Code, to make your setup faster, your systems cheaper, and even build new things and automate parts of your business that weren't possible before.
Let's dive into it. So first of all, what is Jev? And I won't go super deep into this, but basically Jev is a new AI model that is quite interesting because it was released by a co -inventor of ChachiBT.
So this person, Diogo, he made this post that already has something like 38 million views. And he says here that Jev is a new type of frontier AI model that is 20 to 200 times faster and 40 to 400 times cheaper. And if you look at the rate card for this model, that is indeed the case.
It is around 24 times cheaper than Haiku and around 230 times cheaper than Fable 5 .1. And a big part of why that is, is because its output tokens, basically its response to you as the user, is always free. And they only charge for the input tokens, which is essentially your prompt.
And it's super cheap. It's only four cents per million tokens. Now, a big part of why it's so fast and so inexpensive is because Jev can only answer in three shapes.
So when you ask it a question, it can either give you a binary response, whether that statement is true or false. It can provide a selection against a menu of options, or it can also give you a response that is based on scale, let's say from zero to ten. And even though this sounds like a big limitation for the model, this is is actually the genius behind why it's so effective in the use cases that we'll go through later.
But before we go to that, one key thing to remember is that Jev is not actually a large language model. And the way that TypeSafe, which is the company behind Jev, talks about this is that they're saying that Jev is the first System 1 model, whereas all the other AI models like Fable, Astra, and the others are what they're calling System 2 models.
Now, in case you've read this book called Thinking Fast and Slow, that system one and system two dichotomy of how humans and people think might be familiar to you. But essentially, the difference between these systems is that system one is all about thinking fast and making snap decisions. And so Jeff, as an AI model, just optimizes against this.
And so it just outputs classifications and can do that really, really fast and really, really cheaply. Whereas LLMs and other general purpose models like Fable or Astra, they can output text and they do that by writing each word. by one but it gives them more flexibility of what it can output obviously but that would have the drawback of these system two models thinking slower versus its system one counterparts and so really if there's one takeaway from all of this i think the best way that you can use jev right now is to combine both combine a system one model like jev with a system two model like the cloud models and so i'll show you some use cases of how you can get started and actually get value from this today but first let's get you set up so that you can actually use jev And as with any other AI model, it's actually available through a variety of platforms.
You can obviously use it by connecting to TypeSafe, who is the company behind Jev. And as per their ex -post at the time of this recording today, they just announced that Jev is now actually available to everyone because previous to this, just hours ago, there used to be a waitlist to access it. So if you go to this URL, you'll be able to sign up there and actually get your API key to connect it to Cloud.
At least when I was testing it personally and throughout the use cases that I'll go through here, I connected to Jev via OpenRouter, which is the service that always gets updated with the newest AI models as they get released. So you can just access that through this URL. And so to set it up with Cloud or any agentic harness that you're using, it's just one prompt away as usual.
And you can just take a screenshot of this if you need a starter prompt to set that up. Or if you want to prompt in the setup guide for everything that I'll cover here in this lesson, I also made this PDF guide, which you can just grab for free below. And you can just send it to your agent for all of the good nuggets that you can pick up in this video.
So once you set that up and you confirm with Claude that you have access to Jev, like what I did here, now we can get to actually using it. And I'll actually talk about this through three levels by which you can use Jev. And by the way, if you want to learn how to build and sell AI systems that businesses actually pay for, then that's pretty much all we do over at the RoboNuggets community, where not only do you get access to the Claude Living Masterclass, which we update every week and takes you from zero to mastery with the latest on AI, but you also get access to our Agents as a Service course, which walks you through how to actually get paid.
for all these AI skills that you are learning. You also get to be part of a genuinely great community of AI builders. In fact, you can see just some of the recent wins our members are getting from the program right here.
So if you want to start earning from AI, then check that just in the pinned comment below. Now back to the video. And the first one is to integrate Jev with your own agentic operating system, basically the way you work with your agents so that you can get faster results and cheaper systems.
So less token burn. Now, because the way we use agents differ depending on the work that we do, I'm sure that you can also find ways to use Jev outside of what I'll talk about. But just to give you an idea, here are two use cases that I am testing out so far using this model.
The first one is around model routing, which is basically letting Jev automate the choice of the model depending on the task that we are giving Claude. And this is important because remember, it is not really practical for you to use Fable all the time because out of all the models, that is the most expensive. Same thing with Opus.
If you just default to Opus every time, then that can also drain your usage quite a lot. And for a lot of tasks, sometimes Sonnet. and haiku which are the cheaper models are actually enough but the problem there is for you to switch through these models and decide the right model for each task that decision usually lies with you as the user and so there wasn't really a quick and cost -effective way for us to automate model routing up until jev And so to set this up, you can just use this prompt for you to get started.
And just to give you a visual demo of the test that I set up, essentially what I asked Cloud to do is to do a comparison of around 12 prompts with Jev and another one where it's running with Fable 5 .1 every time. And you can see here that because of Jev and the fact that it's actually routing to the right model, depending on the task, it actually resulted to 70 % savings because nine out of those 12 tasks never needed the top model anyway.
So that is quite useful, but obviously you have to try it out for the work that you do. do specifically just to see if the output that you are getting is still good enough in exchange for the tokens that you are saving but it's just great that we now have this new class of ai models that can actually do these types of decisions for us now in practice if you're testing this out i do advise you to make a skill command first where you can switch jev off or on for example here in this cloud session you can see i typed in slash jev on and so for this whole session whenever i assign it tasks cloud will now use jev in order to find the right model for that task One example of that is this where I ask it to find the file path where the Jev router script lives.
You can see that for that task, Jev actually assigned a haiku helper, which is the cheapest model to find that file path. which is much more efficient for your token usage. Because if Jev wasn't there routing to Haiku, then we would have used Opus 5 here, which is the default that I'm using for this session.
The second use case is making Cloud become more efficient when finding the right skills. So again, this is just a visual demo of a test that I ran. But essentially what this shows is 14 tests where if you send it a prompt and you ask it to find a specific skill in my workspace, you can see here that Jev takes much less time to find the right skills versus if you just default to something like Opus 5, for example.
And so in total for those 14 tests, Jev was able to find the right skill within five seconds, while Opus 5 took around 30 seconds. And just to show you how Jev was used in this specific use case, basically your input is the task that you are trying to do.
Jev then looks at that and the options that it can choose from would be your skills itself. So at least for me, if you can see, I have something like 145 skills in my workspace. And so because Jev is really quick, it can almost instantly output the right skill from that list, which Claude then loads.
And so if you want to test that out for yourself, then you can just copy this prompt and send it to your agent. Now, beyond just level one of giving you faster and cheaper systems, if we get to level two, this is actually how we use Jev for more business use cases. Because with Jev, you can actually make automations that are almost at lightning speed and doesn't cost as much as the other AI models.
And just to give a visual demo, let's say you have a hundred emails. And the automation that you are building needs to answer a business question, which for this case, we want to know which of these emails are actually leads that we can contact. But obviously it can be others.
Like if these are customer support tickets, then you can triage which ones are needing the most support. But at least for this demo, what we'll show is Jev doing the classification here in this column. And then we'll also use Haiku as well as Fable in order to show the difference between the speed and costs of these models.
So when I click run, these will now show the time it took for these models to class. classify each of these emails in full so let's go ahead and run that and as you can see that took Jeff like no time at all within less than a second it was able to classify all of those leads whether they're warm, whether it's not a lead, whether it's cold, which is much faster and much cheaper versus these other models.
And so when it comes to business automations, that is where Jev really shines. If you have a huge volume of things and there's a business question that's associated to those things, then this is a good candidate for you to use Jev in. So for example, in enterprise, there's a huge industry with regard to detecting invoice fraud.
Spam detection software also has a good use case for this. Community moderation is another. Same when it comes to high volume.
volume requests for any refunds and even classifying your customers if they are churning or not if you're running a subscription software business for example And so that's the pattern that I think would be good for you to think about in your company or in your business. What are the things that you are receiving in volume that you need to classify?
And if you introduce Jev in there, and because it is so quick and it is so cheap, then you'll be able to upgrade your automations to just a few prompts. And finally, we get to level three, which is building apps that have now just become possible and cost -effective because of system one models like Jev. And again, this differs per person, but just to give you an idea of what I immediately use it for.
In our line, work as you might expect i generate a lot of images as well as videos and i put them all here in my os which i name as rubric now because i have hundreds of images on here it's often the case that i need to search for specific images and let's say if i type in cloud in here unfortunately what this will give me are images where the file names contain the word Claude.
So it's sort of like your standard control F. But if we integrate Jev into that image search, and this is just a quick demo so that I can show you side by side. If I type Claude in here, you can see that the file name search here returns only a few results.
But the search powered by Jev actually enables us to search images and videos by meaning instead of just the file name. And so if you have an application where users need to search for things a lot, then Jev might be good to try to see if that is going to improve user experience. for your app.
Another application that I found that is powered by Jev is this one from Kitze who made this app called Unclutter. And basically what it does is it's a Chrome extension where whenever you toggle it on, it basically auto cleans up pages from any elements that are classified as slop. And the one that is doing that classifying is Jev under the hood.
And it basically just looks at all of the elements in a page. gives a quick decision if they are ads, if they are cookie banners, and it just removes all of that when the switch is toggled. And so there you go.
That is what Jev is and a few use cases and ideas for you to take advantage of this new paradigm by which AI models are created and used. I hope that was useful. And as usual, thanks for watching until the end.
And I'm also curious, like what would you use Jev for? Let me know down below and I'll see you all next time. Cheers.
The Hook
The bait, then the rug-pull.
The creator opens with a real-time side-by-side terminal test showing Jev returning structured classifications in 254 milliseconds for $0.00008, while Claude Fable lags behind at a fraction of the speed.
Frameworks
Named ideas worth stealing.
04:05model
The Three Levels of Jev Implementation
Level 1: Integrate to Agentic OS (Model routing & skill picking)
Level 2: Lightning-Speed Business Automations (High-volume classification of emails, tickets, invoices)
Level 3: Net-New Applications (Cost-prohibitive software like semantic media search and ad/slop blockers)
A progressive framework for leveraging fast classification models from internal developer tooling up to user-facing production applications.
Steal forAuditing and upgrading enterprise agentic systems to reduce token burn.
01:21concept
Three Shapes of Jev Answers
Binary (True / False)
Categorical Options (Selection from an explicit menu)
Scale (Numeric score, e.g., 0 to 10)
The architectural constraint of Jev where outputs are strictly restricted to non-prose structured formats, enabling single-pass execution and free output tokens.
Steal forStructuring prompt templates for deterministic routing.
CTA Breakdown
How they asked for the click.
VERBAL ASK
04:10link
“So if you want to start earning from AI, then check that just in the pinned comment below.”
Seamlessly positioned between the initial concept breakdown and the hands-on technical walkthrough, backed by community member revenue screenshots.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Two five-minute configuration changes, a custom output style and an on-demand skill, turn Opus 5's dense jargon and wall-of-text replies into plain, scannable answers.
A three-line prompt that fans Claude out into paired builder and critic sub-agents until every piece clears a stated quality bar — and the one condition that decides whether it helps or hurts.
Anthropic updated its official skill-authoring guide: keep main files under 500 lines, limit references to one level deep, and replace prompt rules with hooks.
A decision-only model built for picking, not writing, paired with Claude Code across nineteen real automations, from spreadsheet tagging to routing which Claude model handles a prompt.
A screen-recorded tour of 25 prompts, sites, and Claude Code skills, sorted easy to advanced, for anyone tired of interfaces that scream default AI design.
00:12
00:16
00:25
00:53
01:02
01:19
01:37
01:46
01:54
02:03
02:12
02:38
02:46
02:55
03:12
03:21
03:30
03:38
03:59
04:04
04:30
04:38
04:44
05:05
05:23
05:32
05:41
05:50
05:59
06:17
06:26
06:35
06:44
06:53
07:02
07:11
07:29
07:38
07:47
07:56
08:05
08:23
08:33
08:42
09:00
09:05
09:18
09:28
09:37
09:46
09:59
10:12
10:19
10:26
10:38
10:56
11:04
11:13
11:18
11:38
12:18
13:30
12:15
14:05
13:32
21:38