The argument in one line.
Routing a customer request through a millisecond-fast classifier before it ever reaches a large language model cuts both response time and paid API calls, because most incoming requests only need a yes/no decision, not a generated reply.
Read if. Skip if.
READ IF YOU ARE…
- You're running AI on a support inbox, contact form, or ticket queue and every message currently goes through a full chat model, even the ones that just need routing.
- You already use Claude Code with MCP connectors and want a concrete pattern for cutting API costs at scale, not just a demo.
- You're building or maintaining a personal AI assistant and want it to feel faster by skipping the big model for simple decisions.
SKIP IF…
- Your request volume is low enough that API cost and latency aren't a real constraint yet.
- You need every response to go through a single reasoning model for consistency or compliance reasons.
The full version, fast.
A classification-only model called Jev can be placed in front of Claude Opus 5.5 to triage incoming requests before any expensive model call happens. In a live 12-message support-desk test, Opus 5.5 alone took 33.9 seconds and made 12 API calls; Opus 5.5 with Jev took 19.7 seconds and made only 4 calls, because Jev resolved the simple yes/no and priority decisions itself and only escalated messages that needed an actual written draft. The build uses Claude Code connected to Zapier's MCP server, which exposes Jev (via TypeSafe AI) as a tool alongside Gmail, Slack, and other business apps, so the same pattern works inside any Claude Code-based assistant, including a personal one like the video's own Jarvis.
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 – 00:46
01 · Cold open: the dashboard already mid-race
The video opens on a live split-screen dashboard already running a comparison: Opus 5.5 alone has processed 9 of 12 requests while the Jev-assisted side is only at 5, and Zubair frames the question the whole video answers, whether Jev can make Opus 5.5 10x faster and cheaper.
00:46 – 01:44
02 · The scenario: a business support desk
He sets up the premise, a support inbox where emails, website form submissions, and complaints all need to be read and routed to the right department, and explains the dashboard is built to simulate exactly that with 12 real request types running through both a solo-Opus lane and a Jev-assisted lane.
01:44 – 02:32
03 · What Jev actually is
Jev is described as a new kind of AI model, not a large language model like Opus, that specializes in fast yes/no, pick-one, or scale decisions instead of generating text, which is what makes it both far cheaper and far faster for the classification step of a workflow.
02:32 – 04:20
04 · Round one: a single checkout complaint
A test message from a business owner reporting a broken checkout is sent through both lanes. Opus 5.5 alone takes 1.9 seconds and makes one API call; Opus 5.5 with Jev finishes in 0.3 seconds because Jev classifies the request (urgent, needs follow-up, no approval needed) without ever calling Opus.
04:20 – 06:04
05 · Round two: 12 requests at once
Zubair runs a full 12-message batch, mixing domain renewals, refund requests, webinar sign-ups, contact form spam, and an access-grant request, to see how the two lanes compare under realistic volume rather than a single clean example.
06:04 – 08:09
06 · Scoreboard: 33.9s/12 calls vs 19.7s/4 calls
Opus 5.5 alone finishes the 12-message batch in 33.9 seconds and makes 12 Opus API calls, one per message. Opus 5.5 with Jev finishes in 19.7 seconds and makes only 4 Opus calls, because Jev resolves the simple classification requests itself and only hands off the ones that need an actual written response.
08:09 – 09:04
07 · When Jev has to hand off to Opus
Using a request for a small-business website quote as the example, Zubair shows Jev correctly deciding the message needs a drafted reply (not urgent, no approval needed) and handing that single step to Opus, which writes the draft. The combined lane still finishes faster than Opus working alone because the routing decision itself cost almost nothing.
09:04 – 09:39
08 · Where this applies beyond support tickets
He generalizes the pattern beyond customer support, noting any workflow with a mix of simple routing decisions and occasional generated responses benefits from putting a fast classifier in front of the expensive model, and previews the free prompt pack for building it.
09:39 – 11:11
09 · Opening Claude Code and picking the model
Switching to a build walkthrough, Zubair opens the Claude Code desktop app, selects Opus 5.5 with max thinking mode for its price-to-capability ratio, and explains the build needs to connect to whatever tools power a real support desk, such as email and a helpdesk system.
11:11 – 13:41
10 · Wiring Jev into Zapier's Claude MCP Server
Since Jev isn't a native Claude Code connector, Zubair routes through Zapier's MCP server, adding the TypeSafe Jev app alongside Gmail, Slack, and Canva, then walks through creating a TypeSafe AI account, generating an API key, and connecting it so Claude Code can call Jev as a tool.
13:41 – 15:25
11 · Same trick, running inside Jarvis
He demonstrates the pattern already live inside his own personal AI assistant, Jarvis, which also runs on Opus 5.5: Jev classifies each spoken question first (is it a vault lookup, a direct request, just narration) before deciding whether Jarvis's larger model needs to engage at all.
15:25 – 16:42
12 · Grabbing the free prompt pack
Zubair points viewers to a free community where the full prompt pack, and a paid version of the Jarvis assistant itself, are available, walking through the exact click path from the link in the description to the classroom resource.
16:42 – 17:30
13 · Sign-off
He closes with a brief recap that the combination is going to change how he builds going forward, and asks viewers to like, subscribe, and comment with questions.
Lines worth screenshotting.
- Most incoming support requests are actually simple decisions, not response-generation tasks, and routing them to a decision-only model skips the expensive step entirely.
- On a 12-message batch, adding a fast classifier in front of Opus 5.5 cut total processing time from 33.9 seconds to 19.7 seconds and Opus API calls from 12 to 4.
- A single classification request that took Opus 5.5 1.85 seconds took the decision-only model 293 milliseconds, because it never generates a text response.
- The decision-only model doesn't replace the large language model, it decides whether the large model needs to be called at all.
- When a request does need a written reply, routing through the classifier first still finishes faster overall than sending everything straight to the large model, because the classifier's decision step is nearly free.
- Zapier's MCP server can expose a specialized model as just another connected app, letting Claude Code call it the same way it calls Gmail or Slack.
- The same triage pattern works inside a personal AI assistant, not just a business support desk, cutting how often the assistant's main model has to engage for simple requests.
Route the decision before you pay for the response
WHAT TO LEARN
Most inbound requests only need a routing decision, and putting a near-instant classifier in front of your large language model cuts both cost and latency without touching response quality.
02The scenario: a business support desk
- A real support inbox mixes simple routing requests with ones that genuinely need a written reply, and treating every message the same way wastes the expensive model on trivial decisions.
03What Jev actually is
- A decision-only model that answers yes/no, pick-one, or scale questions is a different tool than a chat model, not a smaller version of one, because it never produces free text.
04Round one: a single checkout complaint
- A single classification (urgent, needs follow-up, priority level) took a decision-only model 293 milliseconds versus 1.85 seconds for a full chat model doing the same reasoning step.
06Scoreboard: 33.9s/12 calls vs 19.7s/4 calls
- On identical 12-message batches, adding a classifier cut total time from 33.9 to 19.7 seconds and cut paid model calls from 12 to 4, an 8-call reduction from routing alone.
- Scaled to thousands of daily requests, the savings compound in both time and API spend, since most of that volume is routing, not writing.
07When Jev has to hand off to Opus
- When a message genuinely needs a drafted response, the classifier still adds value by making the yes/no routing decision (urgent, needs approval) before the large model starts writing, so the combined path stays faster than the large model doing both jobs.
09Opening Claude Code and picking the model
- Picking the model with the best capability-to-cost ratio for the reasoning step still matters even after adding a cheap triage layer in front of it.
10Wiring Jev into Zapier's Claude MCP Server
- A workflow automation platform's MCP server can expose a specialized non-chat model as a callable tool, meaning you don't need a native integration to add a classifier to an existing agent stack.
11Same trick, running inside Jarvis
- The classify-then-generate pattern isn't specific to customer support, it applies to any assistant where most interactions are simple lookups or commands and only some require the main model's full reasoning.
Terms worth knowing.
- Jev
- A decision-only AI model from TypeSafe AI that answers yes/no, pick-one, or scale questions about a piece of text in milliseconds, but never generates written output.
- MCP (Model Context Protocol)
- A standard that lets an AI coding tool like Claude Code connect to external tools and services, such as Zapier, Gmail, or a specialized model, and call them as part of its workflow.
- Zapier MCP server
- A hosted connector that exposes thousands of apps, including Jev, as callable tools inside Claude Code or other MCP-compatible clients.
- Opus call
- One billed API request to Claude Opus. Fewer Opus calls for the same volume of requests means lower cost, since the classification step is handled elsewhere.
Things they pointed at.
11:05toolClaude Code (desktop app)
11:05toolClaude Opus 5.5
13:00toolZapier MCP server
15:42linkFree Skool community + prompt pack
16:30productJarvis (paid AI assistant)
Lines you could clip.
00:00
“Can Jev make Cloud Opus 5.5 10 times faster and cheaper?”
05:53
“Jev does not provide any output. It's incredible at taking a set of information with a bunch of different choices and making a quick decision.”
09:53
“This took 33.9 seconds processing 12 requests, and it made 12 Opus calls. This one took 19.7 seconds and processed 12 requests, same thing, but only made four Opus calls.”
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.
Okay, so let's run this. Oh wow, look at how fast the right -hand side is with Jeff and Opus 5 .5 working together versus Opus 5 .5 alone. It already processed nine requests and this side is only at five right now.
This is crazy. Can Jeff make Cloud Opus 5 .5 10 times faster and cheaper? To find out, I built this real -life customer request dashboard.
We're putting 12 actual requests through two versions. On the left -hand side, you got Opus 5 .5 handling everything by itself. And then on the right -hand side, you got Jev and Opus working together.
We'll compare the speed, the number of cloud calls, and how each version handles the request. Then I'm going to explain how this could fit into a real business workflow and show you step -by -step how to build something like this and give you all of the prompts completely for free. follow along and create your own version for your business or a client.
So imagine this is your business support desk, right? You've got emails coming in into inbox. You got people filling in requests on your website.
You got customer complaints coming in and everything needs to be processed and it needs to go to the right department for. that particular request to be handled. So on the left -hand side, we're going to use Cloud Opus 5 .5 alone, which is an absolutely remarkable model that recently got released.
As far as the combination of the smartness versus the efficiency and the amount of money it costs, it's way better than any other model in the market. And then on the right -hand side, you have Jev working with Opus 5 .5 together. Now, if you're not familiar with Jev, Jev is essentially a new AI technology, a new AI model.
It's not a large language. model so it's not like cloud opus 5 .5 it basically is good at making quick decisions and therefore it's incredibly fast and incredibly cheap so in this particular scenario we're going to test a bunch of different things first obviously we're going to go through one or two customer requests process that so you can see in life see which one does a better job and then go through the details and then afterwards i'm going to test actually 12 message sets all at once so we can see that process together and go through the entire details to see whether adding Jeff to something like Opus 5 .5 will actually make a difference.
All right, so the first test we're going to run is a single request item, right? So this is a checkout problem. Let's say somebody named Priya Nair, owner of Juniper Candles Company, right?
Says that, hey, our checkout stopped accepting payments this morning. Customers are waiting, right? Let's say you have a business that is in the processing of payments and somebody sends a complaint like this.
Now, this is very common, right? You can apply this to any sort of business there. So let's go ahead and run the comparison and you'll see how both of these sides handle this So I'm going to click on run comparison Look at how fast Jev with Opus 5 .5 is, right?
It already processed everything in 0 .3 seconds, while Opus alone took 1 .9 seconds. So now let's go ahead and compare actually what happens. If I click on this, all right, so this is essentially a classification problem, right?
A decision needs to be made whether something like this is urgent and whether this requires a follow -up, right? That's how simple that task is. Now, as far as the processing, of course, Opus is gonna take its time, right?
Because it needs to actually grab, that information, process it, and then send a request outside or send an output. On the right hand side, you see Jev actually did this by itself.
It didn't even call Opus 5 .5 because in this particular scenario, it didn't need to, right? So this is a simple classification and needs a quick decision. That's where Jev is really good at because there is no output.
Jev does not provide any output. It's incredible at taking a set of information with a bunch of different choices and making a quick decision. That's exactly why this took 293 milliseconds versus 1 .85 seconds that cloud opus 5 .5 took right so all it did was it made the decision that okay this is a follow -up is it urgent yes it is needs approval no tasks saved and it said moved it to priority number one right that's how simple it made this request instead of having Claude go through the whole process step by step and making that decision and now you can apply this in a scale so let's say you've got thousands of this type of problem that are coming in and it just needs you know classification like this then you can imagine how much time money and most importantly how much API requests let's say if you're using Opus in the API which most businesses do is going to save you Now let's go ahead and try, let's say, 12 message set, right?
Because this will be the ultimate test so you can see exactly how long both sides take. And then we're gonna go ahead and take a look at the results. And like I said, then I'm gonna talk about why this could be applied to a business and show you guys how to build something like this or a version of this.
Okay, so I'm gonna click on the 12 message set. So you can see it says 12 messages, same order in both lanes. Your domain renewal receipt, new menu for website, refund request, webinar, contact form sends, blank emails, give Jordan admin access.
side down quick one right so there's lots of different variation of the requests that are coming in not just a simple classification where we just went through with this checkout problem right so now you'll see how claude comes into picture here when you're using it together with jeff so let's go ahead and click on run comparison look at that already on the right hand side you can see it processed eight requests in under three seconds and claude is still on request three process nine requests right so you can see some of these require jeff to work with claude together or with opus 5 .5 together that's why this is taking long versus the one that's just very quickly just jeff is doing everything by itself right jeff is already done and opus 5 .5 is still processing this so that's how big of a difference this makes and like i said you can apply this at a scale and you can see how much time and how much more efficient and cost effective this becomes when you're combining something like jeff with a powerful model like opus 5 .5 all right so they got done so this took 33 .9 seconds processing 12 requests and it made 12 opus calls right
This one took 19 .7 seconds and processed 12 requests, same thing, but only made four Opus calls. So if you look on top right here, the first few, right? So on the left hand side, obviously, you're going to see only Claude because Opus is doing everything by itself.
On the right hand side with these combinations, as you can see, the first few items did not even require Claude to jump in or Opus to jump in. That's why Jeff did all of it by itself. Right.
So and then on the these are the ones that require Jeff to essentially hand it over to Claude. That's why you see that the request between. jeff by itself versus with opus 5 .5 together takes this one takes 3 .57 seconds that takes 246 milliseconds so if we compare this so this one let's say price for small site if we go to that one right price for small site This requires an actual draft or an actual response, right?
So that's why Claude is needed here, obviously. So if we click on this, price for a small site. So somebody sent a request saying, I run a plumbing business and need a simple website, maybe five pages.
What would that cost and how long does that take, right? So now this is something that's going to require a response back. Therefore, this...
jeff on the right hand side said okay this one actually needs a draft reply is it urgent no it's not does it need approval no so that's where now opus wrote the draft so this one jeff said okay i've decided that this is not something that i could do myself obviously right because this requires a response therefore i'm going to get opus involved so that's why the draft for both of them looks exactly the same right because same thing on the left hand side opus did this draft automatically on the right hand side the save draft was written but But of course, like I said, that quick decision was made by Jeff.
Therefore, this side is still way faster than your Opus. So this side, it took three, three and a half seconds versus this one where it took almost five seconds. So even when Opus gets involved, you can see that this side is a lot faster and a lot more efficient.
Now, same thing you can compare on the bottom. So you can see adding an online store maintenance after lunch. So these are the ones that require draft or some kind of processing.
Therefore, Claude gets involved and Jeff makes that request. there so in the bottom i'm just you know uh filtering this with with with different scenarios whether something requires a task a draft or a review and uh just kind of comparing the two side by side but that will give you that essentially gives you a good idea of how good and more efficient this is now of course this could be applied to any business scenario right you don't have to just look at it from a ai customer support perspective this could be applied to any scenario but ai customer support is probably the best situation for this so i can so that way we can actually see this and work so if you want to build something like this i'm going to show you how you can gain access to the prompt pack in a little bit again it's going to be completely free i want to put the link in the description but i'm going to show you exactly how to get that because sometimes people get confused there but of course we're gonna build all of this you would have to build this inside clot code so i'm going to show you the few tools that you need to attach in order to build something like this so that way you can just follow along using that prompt pack
All right, so I'm inside my Cloud Code. Of course, we're gonna use the desktop app. So if you don't have that installed, please make sure you install that because that's gonna make your life a lot easier, right?
So after you're here, just go to Cloud Code. What we need to do is, of course, make sure you're using Opus 5 .5 just because it's the best model right now, in my opinion, of how efficient it is, how fast it is, how smart it is, right? It's basically a Fable 5 .5, 5 .1 with half the price.
Because the input token and output tokens are a lot cheaper So once you select that I'm keeping it at max, but feel free to use extra as well That works really well as well. Alright, so in order to build something like that what you need to do is first of all connect this to an mcp tool that could give you access to all of the tools that you're using let's say for your customer support right so let's say you have a customer business that or if you're building something like this dashboard that we looked at of course there's going to be lots of tools that needs to be connected together if this is going to be a real life scenario so a lot of the requests that are inside this some of them came from email some of them came from a contact form that was sent on our website and then a bunch of other ones that was coming in from different resources and of course for something like this it needs to connect to different tools like zendesk or any other support tool so that's why we need to be able to in order to do that we need to have clot code have the ability to access all those tools now in order to do that the great thing is clot supports connectors so you can just go to connectors right you can go to browse connectors
And you can look for individual, those tools that you're using, right? And unfortunately, Jev is not part of Cloud Code right now. But in order to connect all of them, there's a shortcut, right?
So what I'm using, what I use to build this, and of course, for anybody, I would recommend using something like a Zapier because that will connect everything together. And you can add Jev directly there. So if you just search for Zapier.
There you go. So it's this one so if you click on this I'm gonna put the link in the description make sure you click on that link because the first thing you need to do is sign up for a Zapier account and it's gonna bring you to this MCP page right here if you click on the link It's gonna bring you directly right here And as you can see I have my type save jab already connected here and all other tools that's gonna require Something like this, right?
So Gmail slack Canva or whatever and then in your situation if it's coming into a website request or if you need to send that and follow that to your customer support team, so you're going to need something like a Zendesk or whatever else you're using, right? And it's very easy to look for these things.
So there you go. All you have to do is just look up that app, click on this, and you just basically connect exactly the permission that you want to give. As far as the tool itself to connect Jev, so all you have to do is search for Jev here, and then this one is called TypeSave Jev.
You're gonna click on this. I've already added mine, right? But for yours is gonna be let me actually remove this I can show you guys how to install this from start.
So I'm gonna get rid of this Remove tool. Alright, so now if I click on add apps and look for Jeff Click on type save Jeff. There you go.
So now I can click on this, right? It says ask questions because it says ask yes or no, pick one or scale questions about some text and returns each answer. And that's exactly what we want, right?
Because that's what Jev is good at. And that's what that process of first initial step, which is classification, happens all through Jev here. So once you click on this, you're going to click on connect.
And this is going to open up this page where it's going to tell you to connect your Jev account. Of course, mine is already connected, but yours is going to look something like this. All you have to do is click on connect.
I'm going to do connect new account. So this is going to open up this connection where you would have to just attach your API key. Now, in order to get your API key, it's very simple.
You're just going to head over to typesave .ai, sign in. After you sign into your account, if you're in a homepage, it's going to look something like this. You're going to get to the API page.
You're going to click on Create Key. Just name it something, right? Oh, that's not what I meant.
Right? So there, Create Key. And now you can just copy it.
By the way, I'm gonna delete this API key. So don't even think about using it You're gonna paste it click on connect And done right so now I have my Accounts connected here. So both of these accounts are mine, of course, but that one uses my previous API So now I can choose my account here and you're gonna click on add to and now you have your tool added on your mcp server here so of course you want to make sure that whatever if you're using let's say gpt codex right so you can do that or if you want to create something custom that's exactly what you can do so for instance for my jarvis server right so if i uh show you right here so this is jarvis i also use jeff inside jarvis so if i ask him a question jeff classifies that question and then goes and uses the opus 5 .5 brain which is also inside jarvis let me show you an example Jarvis, can you pull up my free school community?
Look at that. Jev is doing the classification and this is what... I've pulled up all the information, sir.
You can read it yourself or if you'd like, say, read it out loud and I shall. All right. Thanks Jarvis.
So it already, Jeff classifies the question. Hey, is this a vault question? Have I completed talking the request?
Am I addressing Jarvis straight or am I just narrating? And is this something that's going to require some kind of a tool use? And is this something that is inside my vault, which is this AI second brain, right?
So this is my free school community and all of my business related stuff is right here, right? So that's UCF representation of this. And Jarvis is my AI assistant that handles everything.
thing and therefore like i said This just makes something like Jarvis, which again runs on Opus 5 .5, even more efficient because it's classifying my question and then routing it to make sure that Jarvis is not processing everything using its big brain, which is Opus 5 .5. So that's how you can use this practically in all sorts of application.
And same thing, like I said, for a real business, obviously, like I said, you have a bunch of different tools that are working together. And that's where you can now utilize and connect all of the... tools that you're using and then of course use cloud code with this connection right so that connection is going to look something like this and if we go back click on the plus button the connector you want to make sure that your zapiers is activated and then now you can use so uh the prompt guide so let me pull that up for you guys so this is the prompt pack right that's going to explain everything what you're building and the different instructions that you have and then also give you the prompt pack so that way you can build something like a customer support dashboard that we looked at or you can apply this to whatever you're trying to build with cloud code using the MCP connection there.
So in order to gain access to that, all you have to do is click on the link in the description. It's going to say free community, free prompt pack. You're going to come to the community here.
You're going to click go because it's going to bring you to the about page. Click on join. Again, this one's completely free.
You're going to go to the classroom section. go to YouTube resources, and then you will see this inside Cloud Code and Codex. Just go to the bottom.
You'll find the prompt pack that you can download very easily and then just follow all of the directions there and you'll be able to build something similar to this or whatever it is for your business. And if you're interested in grabbing a Jarvis, a full version of Jarvis, the AI assistant here, that's going to be also in the link in the description.
That's going to be inside the paid community, right? So inside the classroom section. AI assistance in AI employees.
You just download this really quickly, the zip file, upload it, and you'll be able to grab your Jarvis that we just saw here for your own self. So that way you can utilize that. Again, all of these links are going to be in the description.
I don't want to make this video too long because this is already gone very, very long. But again, this is absolutely incredible. I'm going to try to build more things here with this particular combination because I think this is going to definitely change the way we build things and it's going to make everything a lot better, faster.
more efficient and cheaper. Anyways, hopefully you guys found this video helpful. Make sure you like, subscribe so that way you can see the upcoming videos about topics like this.
Let me know if you have any questions in the comments below. Thanks for watching and I'll see you on the next one.
The Hook
The bait, then the rug-pull.
Zubair opens mid-test: a support-desk dashboard already running, one column labeled Opus 5.5 alone, the other Opus 5.5 paired with a model called Jev, and the Jev column is visibly pulling ahead. The rest of the video is him explaining exactly why.
Named ideas worth stealing.
02:32model
Classify-then-generate routing
- Fast classifier decides urgency/approval/routing
- Only messages needing a written response get escalated to the large model
- Large model is called once per escalation, not once per message
Every incoming request first passes through Jev for a cheap, near-instant classification decision. Only the subset that actually requires generated text (a drafted reply, a written response) gets escalated to Claude Opus 5.5.
Steal forany inbox, ticket queue, or webhook pipeline where most messages need routing, not writing
How they asked for the click.
VERBAL ASK
15:25link
“Click on the link in the description. It's going to say free community, free prompt pack.”
Soft CTA delivered after the value is fully shown, pointing to a free Skool community with the build's prompt pack, and mentioning a separate paid tier for his full Jarvis assistant.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
