21:08Riley Brown · Tutorial
Building a Real AI Jarvis with Cursor and GPT Realtime 2
A step-by-step build of a talking, screen-controlling desktop AI companion — no coding experience required, just long prompts into Cursor.
July 1stThe argument in one line.
Grok 4.5 performs close to Claude's Opus-tier models on real coding tasks while costing roughly a quarter as much and responding faster, and pairing it with Cursor's Canvas and Design Mode turns app prototyping into a near-instant, iterative process.
Read if. Skip if.
READ IF YOU ARE…
- You build with AI coding agents and want to know whether a cheaper model can replace Opus-tier output for everyday app and site work.
- You're deciding which model to route through Cursor and want real cost and speed numbers instead of benchmark scores.
- You want practical examples of Cursor's Canvas and Design Mode features used to prototype a website, a mobile app, and a web app in one sitting.
SKIP IF…
- You need rigorous, independently-scored benchmark comparisons — this is one creator's hands-on impression, not a formal eval.
- You're not using Cursor or a similar agentic coding tool, since the specific workflow demoed depends on that environment.
The full version, fast.
Grok 4.5, xAI's new model now shipping inside Cursor, gets put through four builds: a pricing-comparison landing page made with Cursor's Canvas feature, a personal website revamp, an iOS voice-chat app built in Swift and run in the Xcode simulator, and an Excalidraw clone backed by a Convex database and deployed live to Vercel. Combined input-plus-output pricing puts Grok 4.5 at about $8 per million tokens versus roughly $60 for Claude's flagship and $30 for GPT-5.6 — close to four times cheaper than Opus-tier models while matching much of their coding quality. The takeaways: it's unusually fast, it pairs well with Cursor's Design Mode for rapid visual iteration, and outside builders (including Elon Musk) are calling it a legitimate daily driver rather than a benchmark trick.
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 – 00:21
01 · Intro
Cold open: Grok 4.5 ships inside Cursor, pitched as near-Opus quality at a fraction of the cost.
00:21 – 01:47
02 · Grok 4.5 for knowledge work (Canvas)
A single prompt generates a full model-comparison landing page with charts and quotes inside Cursor's Canvas feature; iterated with follow-up prompts (recenter, add charts).
01:47 – 02:58
03 · Grok 4.5 pricing
Combined input+output pricing chart: Claude's flagship at ~$60/M tokens, GPT-5.6 in the middle, Grok 4.5 at $8/M — reads Elon Musk's public statement on cost vs. benchmarks.
02:58 – 03:26
04 · Prompting for a landing page
Prompts Grok 4.5 to revamp rileybrown.xyz using Cursor's own landing page as a style reference, optimized for mobile and web.
03:26 – 07:15
05 · Creating a voice-mode Grok app
Opens a second workspace and builds an iOS Swift app that lets you talk to a Grok voice agent, tested live in the Xcode simulator; app works but has audio/duplicate-message bugs.
07:15 – 08:38
06 · How is Grok 4.5 at design?
Discovers the earlier landing-page prompt was pushed live to the real site instead of run locally; reviews the result in both dark and light mode.
08:38 – 10:20
07 · Design Mode with Grok 4.5
Runs the site locally and uses Design Mode to click-fix individual elements — resize, visibility, color, footer copy, view-count formatting — in rapid single-element passes.
10:20 – 12:00
08 · Excalidraw clone with Convex
Prompts a from-scratch Excalidraw clone with a real Convex database and no auth; Grok 4.5 deploys it live to Vercel unprompted within about five minutes.
12:00 – 15:25
09 · What other people are saying
Recaps the three builds, then reads reactions from Elon Musk, Cursor's CEO, Theo, Dax (OpenCode), and Evan Bacon; notes the coming context-window bump and a much larger model in training.
Lines worth screenshotting.
- Grok 4.5 costs about $8 per million combined tokens versus roughly $60 for Claude's flagship model — nearly four times cheaper for comparable coding output.
- Cursor's Canvas feature generated a full comparison landing page, complete with charts and pulled quotes, from a single natural-language prompt in under a minute.
- An iOS voice-chat app built entirely by prompting Grok 4.5 in Swift was running in the Xcode simulator in about three and a half minutes.
- Cursor's Design Mode lets you click a specific UI element — a chart, a footer, a tag — and describe the fix in place, instead of writing a broader prompt.
- Grok 4.5 defaulted to deploying an Excalidraw clone live on Vercel with a Convex database, without being asked to skip a local/staging step.
- Elon Musk says Grok 4.5 is roughly comparable to Claude Opus but faster, framing cost and speed — not benchmark scores — as the real differentiator.
- A developer from OpenCode reported using Grok 4.5 as his daily driver the day after release, calling it the first xAI model that clears the bar for everyday coding work.
- Grok 4.5's context window is expected to expand to 1,000,000 tokens within about a week of launch, according to Elon Musk.
- xAI is reportedly training a 2-trillion-parameter follow-up model.
- An agent instructed to 'revamp' a live personal site pushed the change straight to production instead of running it locally first — a reminder to specify the target environment explicitly.
Cheap and fast can beat premium when picking a coding model.
MODEL SELECTION
Grok 4.5 shows that raw benchmark superiority matters less than combined speed and cost once you're iterating on real projects all day.
02Grok 4.5 for knowledge work (Canvas)
- A single detailed prompt — audience, data points, tone, and format — can produce a usable interactive artifact instead of a static answer.
- Iterating after generation, like recentering a layout or requesting more charts in a matching style, often takes less time than getting the first prompt perfect.
03Grok 4.5 pricing
- Comparing frontier models on combined input-plus-output price per million tokens gives a clearer cost picture than looking at either number alone.
- A roughly 4x price gap between a premium model and a competitor is large enough to change which model becomes your daily default.
- Public statements from a model's own creator about being 'comparable but faster' should be treated as a hypothesis to test, not a verdict.
04Prompting for a landing page
- Naming a specific reference example, like a competitor's color palette without copying it directly, gives a coding agent a usable style anchor without dictating exact output.
- Running multiple independent workspaces in parallel lets one long task keep moving while you start a second, unrelated build.
05Creating a voice-mode Grok app
- Testing in a simulator instead of a physical device removes a dependency and speeds up the iterate-and-check loop.
- A capped-spend API key is a simple guardrail when handing credentials to an agent you don't fully trust yet.
- Fast generation doesn't guarantee correctness — a working demo can still ship with real bugs that need a follow-up fix prompt.
06How is Grok 4.5 at design?
- An agent can interpret 'make this better' broadly enough to push directly to production instead of running locally — confirm the target environment before firing off a build prompt.
- Explicitly stating 'run locally, don't push yet' is a low-cost habit that prevents unintended live deployments.
07Design Mode with Grok 4.5
- Selecting a specific UI element and describing the fix in isolation gets faster, more targeted results than one broad rewrite prompt.
- Small, sequential design tweaks compound into a materially different final layout within minutes.
08Excalidraw clone with Convex
- Stating a constraint like 'no authentication' up front is a legitimate way to scope a build down to just the core feature you're testing.
- A model that defaults to a real database and a live deployment, rather than a local mock, is optimizing for a working product rather than just a demo.
09What other people are saying
- Weighing reactions from people whose judgment you trust is a reasonable substitute for reading every benchmark chart yourself.
- A model's context window and parameter count keep expanding after launch, so day-one impressions can shift within weeks.
Terms worth knowing.
- Cursor
- An AI-powered code editor that lets you prompt for app changes directly, with features like Canvas (interactive artifact generation) and Design Mode (click-to-edit UI elements).
- Canvas (Cursor feature)
- An interactive workspace inside Cursor for generating and iterating on visual artifacts, like a landing page or chart set, from a single prompt.
- Design Mode
- A Cursor feature that lets you click a specific on-screen UI element and describe a change to just that element, rather than rewriting a whole page.
- Convex
- A hosted backend and database platform that stores app data without the developer having to set up their own server or database instance.
- Excalidraw
- A web-based whiteboard tool for hand-drawn-style diagrams, sketches, and simple presentations.
Things they pointed at.
00:20productGrok 4.5
02:58linkrileybrown.xyz
10:35toolConvex
11:10toolVercel
13:32channelOpenCode (Dax)
14:40productEvan Bacon rocket-tracking app
Lines you could clip.
00:05
“It's about as powerful as Opus 4.8, but it's faster and cheaper than the Opus model.”
02:17
“It's almost four times cheaper than Opus 4.8.”
02:23
“We are closing the loop on real world usefulness, not benchmarks.”
13:32
“I've been using this as my daily since yesterday... the first model of theirs that clears the bar for day to day work.”
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.
metaphor
00:00Cursor is now part of SpaceX and they just released a brand new model, Grok 4.5, and this model is very good for how cheap it is. It's about as powerful as Opus 4.8, but it's faster and cheaper than the Opus model.
00:15And when you use this model inside Cursor, especially with their new Canvas feature, it's so much more fun to use. Let me show you what I mean. So on Cursor, you can see a new model available called Cursor Grock 4.5 high fast.
00:28And so what I can do is I can say please make a landing page for grok 4.5 which compares the model with the other frontier model companies and make this in a canvas and have really useful graphics and have quotes from what people are saying about this model.
00:53And so Cursor, as you can see here, works incredibly fast and there you go.
00:58That did not take long at all and here we have this clean landing page that we can open up full screen. We could even close this side panel and I could say something like, can you please center this whole page here?
01:14It's left aligned. Please center it.
01:18And then I can also use design mode again. I can select this chart and be like, I love this chart actually.
01:26Can you please make more charts similar to this? Like these colors, I love them.
01:31Add one at the top. I don't know. Make it useful.
01:34And so we can run that. And so it very quickly created this research report directly inside Cursor and I just love this like full screen window. This is how I've been primarily using this new Cursor Grok 4.5 model.
01:47So this is an artifact that I created right before filming and I combined the input and the output token prices to create just one price here. And Fable five is $60 per 1,000,000 tokens input plus output.
02:03And if you take a look here, we see Grok 4.5 is $8 per 1,000,000 tokens input plus output. And if you compare that against Opus 4.8, which many people are saying it's pretty close to as good as Opus 4.8, look at how much cheaper it is.
02:18It's like 3 and a half. It's almost four times cheaper than OPUS 4.8. As Elon Musk said here, said our internal assessments is that Grok 4.5 is roughly comparable to OPUS 4.7 but much faster.
02:30The combination of capability, faster speed, and lower cost is what makes it competitive. We are , not benchmarks.
02:39But I guess we can be the judge of that. So I'm gonna press command n to create a new chat and I'm just working in a folder. Right?
02:45This is very similar to Codex in the sense that I can come up to workspaces and create a new workspace. I just have a general folder, I don't know why I forgot the l, and this is where I just fire off my experimental prompts here.
02:56And so what I'm gonna do is I'm gonna say, hey, can you please look at rileybrown.xyz? That is my personal website. I want you to revamp it.
03:03I want you to make it really good. And actually, I want you to use kind of the cursor landing page as an example, but not too much like cursor, just kind of that that that color palette.
03:13And and I want you to make a new landing page for me that's significantly stronger and better optimized for both mobile and web.
03:24And so we can just run this. Now what I wanna do is while that's working, let's go ahead and create a mobile app. And instead of working in the same workspace, I'm gonna go ahead and create a new workspace.
03:34We're gonna come down to open folder, and what I'm going to do is I'm going to create a new folder in my documents.
03:43We're gonna go mobile app testing Grok.
03:48And what I'm gonna do is I'm going to create a mobile app here. And what we're gonna do is we're gonna build an iOS app in Swift that is the Grok voice agent. We're gonna use this Grok voice agents API, and all I need to do is go up here and grab an API key, which I've already done.
04:04And what I'm gonna do is I'm gonna go back to cursor, and I'm just gonna say the following. We've created a new folder. Mhmm.
04:11It's called mobile app testing Grok. I want you to create a mobile app and I want this in Swift. I want to create the mobile app that it has a really cool design, black background, cool design that allows me to chat with one of the voice agents by Grok.
04:29Here is a link to the documentation. Let's go ahead and paste this in here. I'm going to run my prompt and then in the next prompt I'm gonna upload my API key.
04:40I want you to use this to create a really high quality Swift app. I want you to run this in the simulator, so I wanna be able to test it using my computer. I don't wanna have to use my phone.
04:49We're gonna get this running in Xcode and run it in the simulator. And so I'm gonna go ahead and fire off this longer harder prompt and then I'm gonna paste my API key. I know that they say not to do this, but I do it anyway.
05:01I'm just gonna paste my API key in the chat here. Okay. This model is really fast.
05:06Oh my god. This is already done. What the hell?
05:10Okay. Don't worry. I'm gonna cancel this key, and I only have $10 on my account, so that's okay.
05:16But take a look at this. It worked for a total of three and a half minutes, and it's already done.
05:23Now I have no idea how good the app's gonna be. It says it's running. Okay.
05:27We got it running in the simulator. Let's see if this works. Wait.
05:31How does how does this work? Do I play?
05:37Hello? Hello there. I'm so glad you're here.
05:39I'm Grok, your warm and witty companion. Are you able to hear me? I can't hear you directly.
05:46I'm text based, but I'm all ears for whatever you type or say through this chat. Okay. Just let me know what's on your mind, and I'll this?
05:54Like, it the the my audio I'm not able to interrupt the AI model, and, like, things are coming through, like, in a patchy way. Can you please just clean this up? Like, really think deeply about this and fix this?
06:08Because are you still listening to me?
06:12I'm still right here with you listening through f Okay. So it's kind of working,
06:17but it's sending my messages in twice. But it still created this iOS app, and I can chat with Grock, which is pretty cool.
06:25Okay. So I actually have it starting up on my Nice to meet you.
06:29How's it going? Hey. Are you able to hear me right now?
06:33I just want you to tell me that you love me.
06:38I can hear you loud and clear and I love you. Nice.
06:43Nice. Nice. What's the best way to make some friends?
06:48The best way is to put yourself out there and join activities you enjoy. Be friendly, start conversations and follow-up if you hit it off. Okay.
06:57So I think Grok is actually pretty good at creating iOS apps. It made one little mistake with like, I think it's like the web sockets or something with the voice but this is pretty good and it generated it really fast. Now what I wanna do is I wanna take a look at my landing page that I created a few minutes ago.
07:15Oh my god. Okay. So I asked it to change my landing page and I was hoping it would just scrape my landing page, make a new one and run it locally.
07:25But it turns out that it actually pushed it live to my actual personal website. So let's see what this looks like here. So I'm gonna go ahead and I'm gonna open this up.
07:35Okay. Okay. Let's see.
07:37Let's see Grok. See if it cooked. Okay.
07:39So this already looks a little bit better. I wouldn't choose this photo. This is just my Twitter profile photo, which I need to update.
07:45But this looks pretty good and it has all of my live stats here. Let's see where to go to next.
07:52Very cool. Uh, recent long form.
07:59Solid. This is a cool website. Let's see what it looks like on light mode.
08:02Not bad.
08:05What I like to do with cursor is I actually like to run it in the in app browser. So what we're gonna do is I'm gonna say run.
08:15Can you please run this locally and not push to my site yet?
08:24I want to to make some changes. Okay.
08:29So now that it's running locally, what I like to do is full screen this and what I'm gonna do is I'm gonna close this left side panel so we kinda have it full screen. And to bring up the chat while in full screen mode, you can just work on it from here.
08:42You can choose any model. Right now, we're using the new Grok 4.5 high fast. And now, we can just use design mode.
08:49And so, I'm just gonna go through this really quick. This is slightly too large.
08:59And we can just kind of like go through all of this.
09:06We don't need the add tag. Make this more visible.
09:13Make it a more purple color.
09:23This is pretty cool. What we can do is we can just click on this. Please make the footer better and just use your best judgment to, like, what would the best personal website footer here be?
09:32Maybe references to other links that we have on the site? And made my social links again here. Okay.
09:39So it's just generating it out, which is sick. Oh, that looks good. So this model's fast, dude.
09:46Alright. So it made that change. It changed this.
09:50Can you please make change this to like a view icon instead of like views written out and make it like 242 k instead of listing the exact amount of views.
10:06Yeah. And this is fun. We're just firing prompts off to Grok and it's making the changes.
10:14Okay. This is a model that just works every ten seconds. You know, this worked for nineteen seconds.
10:20Okay. So now let's try something, uh, a web app that's a little bit more challenging. Just to continue to test this new model, I'm gonna go ahead and open a new folder here.
10:28We're gonna go desktop and I'm going to hit, um, new folder. We're gonna call this create open.
10:36Hey. I want you to please create a perfect Excalidraw clone except I want to be able to turn them into little presentations.
10:45Actually, you know what? I want this to be, yeah, I want this to be a web app and I want all of the data to be stored well. So make a database.
10:54No authentication. It's a personal website for me, but I want the data to be perfect and, um, all of it should be stored perfectly. Make it a perfect clone.
11:03Okay. So now it's going to work. Let's see how Grok 4.5 does here.
11:07God, it just works fast. Okay. It looks like it's going with Convex and it's gonna just deploy it to prod with Vercel.
11:17Okay. Very interesting. Wow.
11:19It's already deploying. It's only been going for like five minutes. Let's see this deployed app.
11:26This is actually on the Internet. You can go to this site and use it. Welcome to Excal.
11:32Okay. New drawing. Hello there.
11:35My name is Riley. Let's see if we can do all the things.
11:40We can change the colors. Let's see if these work. Hello there.
11:45Very cool. We can press 5. We can use the arrow.
11:49Okay. This is cool. Created by Grok.
11:56And this is deployed on the Internet. So this is an actual link that is running Excalidraw.
12:04Here it says five elements. Let's see if I can rename it. Grok made this.
12:11Very cool. Awesome. Okay.
12:13So before we get to what other people are saying about the model, let's talk about what we've done. So we created a website. This was a personal landing page, a rendition of my current one, and it literally just deployed the new one straight to Vercel, which was interesting, but I think it did a pretty good job designing.
12:28It also created this voice mobile app where you can chat with Grok live voice, and it also created an Excalidraw clone, and it did a pretty good job. And so my main observations are just that it's really, really fast. It's four times cheaper than Opus.
12:42It's also really fun to be to use it with the cursor design mode. It's very easy to just, like, really quickly iterate. And it's also really good with the cursor canvas, I showed you that at the beginning.
12:52It can create these little artifacts that can be shared with your team, which is really fun. So the first thing that Elon is iterating along with Michael Troule, is the CEO of Cursor, and remember SpaceX now owns Cursor, he says that it is an Opus class model that's fast and low cost and it's a significant step up over models we've developed so far including Composer 2.5 and has become the .
13:16First of many releases more soon. And many people who I respect on Twitter has been saying that this model's really good and that they're actually really enjoying it. Theo said that Grok 4.5 is good enough that I think I'll have to do a video on it.
13:31And then Dax from OpenCode said that I've been using this as my daily since yesterday. It's super fast and I think it's the first model of theirs that . Also worked well with browser control and took care of some admin work for me.
13:45And so this is one thing I didn't test that I may test more in the future which is using it for knowledge work And so I'll definitely have to test that one out. And if you watch my channel, you know that I don't really pay attention to benchmarks that much. I'd rather judge my own intuition or talk to people that I trust to see how they like the model.
14:02And but it is scoring pretty well. But the interesting thing to note here is that Elon Musk said that Grok four point five's context window will upgrade to 1,000,000 probably by next week. So that means this model is gonna get better very quickly.
14:17And they're also releasing more models. Apparently, they're training a 2,000,000,000,000 parameter model that should be releasing relatively soon.
14:25And honestly, the mobile app that I tried to make probably didn't test its design skills well enough, but this one looks sick. This is from Evan Bacon, um, who just left expo, and he said it just built me this rocket tracking app with live data and a three d globe.
14:41I might need a new benchmark after this. This is really cool actually. Anyway, this is the first model that SpaceX and Cursor have trained together and I highly recommend trying it out.
14:54It's really fast, and there's never been a better time to be interested in AI agents, whether it's for coding or general knowledge work because over the last ten days, we've seen Fable. GPT 5.6 comes out today. If you're watching this on Thursday, I'm filming this on Wednesday.
15:10And now SpaceX just launched their new model. There's so many things happening in the agent world, and I'll be talking about all of the biggest updates that can actually help you use agents in your business. Anyway, thank you guys so much for watching.
15:23I'll see you here for the next one.
The Hook
The bait, then the rug-pull.
xAI's newest model landed inside Cursor this week, and the pitch is blunt: it performs close to Claude's Opus tier at a fraction of the price. Riley Brown spends the next fifteen minutes putting that claim through four real builds — a landing page, a personal site, an iOS app, and a live web tool — to see if the savings hold up once the prompts get complicated.
How they asked for the click.
MENTIONED ON CAMERA
Visual structure at a glance.
