New Gemini 4 Argon vs Opus 5.5 Building the Same App (Raw Results)

Author: Jack Roberts · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: New Gemini 4 Argon vs Opus 5.5 Building the Same App (Raw Results) · Channel: Jack Roberts · 21:06 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, Model inference and optimization, AI models

Summary

Jack Roberts compares Gemini 4 Argon with Opus 5.5 across seven increasingly varied tasks, using the same prompts and tools to build websites, animations, games, and apps. He finds that Gemini is often much cheaper and uses fewer tokens, but Opus generally produces more polished, usable results and wins his overall comparison. The tests also highlight that the development harness can affect output, and that benchmark scores do not necessarily predict performance on real projects. His recommendation is that Gemini is worth considering for cost-efficient work, but Opus 5.5 remains more consistent for these tasks.

Key points

The comparison covers seven tasks, from building an electric-perfume website to designing a thumbnail for the video.
For the website and animated product intro, Jack preferred Opus 5.5’s visual results, although Gemini was cheaper.
Jack says Gemini performed better when used through Google’s Antigravity harness than through Open Router in his tests.
Both models built playable voxel worlds; Jack judged the result close, with a slight edge to Opus.
In the photo editor and personal dashboard tasks, Gemini offered some strengths, but Jack favored Opus’s interface and overall usability.
Both models struggled to impress Jack with the thumbnail and trailer task.
Across the seven tests, Jack says Opus won overall, while noting one result was questionable.
Gemini used about 4 million tokens compared with Opus’s 22 million, but Jack argues that follow-up revisions can reduce the practical value of that efficiency.
Jack’s final view is that Gemini is a capable, cost-efficient model but is not as consistent as Opus 5.5 for the tested work.

Chapters

Start	Chapter	Summary
00:00	The comparison and test setup	Jack introduces the Gemini 4 Argon versus Opus 5.5 comparison, explains how to access Gemini, and begins the first website challenge.
03:01	Website and animated product intro	Jack compares the perfume websites and habit-tracker animations, favoring Opus’s visual polish while noting Gemini’s lower cost.
05:36	Harness explanation and voxel game	Jack explains his use of an API for generated assets, discusses the Antigravity harness, and compares the models’ Minecraft-style worlds.
08:43	Browser photo editor	Both models build a browser-based photo editor, with Jack seeing more functionality in Gemini’s version but preferring Opus’s cleaner interface.
11:49	Personal app and dashboard feature	Jack tests how the models handle a vague personal-app prompt and a feature for his dashboard, and judges Opus easier to use and better designed.
16:25	Thumbnail challenge and overall results	The models attempt a creative thumbnail and trailer task before Jack reviews the overall wins, token costs, and his recommendation.

Quotes

I found the results were so bad. — 05:05

I’d say this was neck and neck. — 08:12

Claude itself won seven out of seven of those results. — 18:30

we spent 4 million tokens on Argon, and Claude spent 22 million tokens — 20:02

it’s not quite as consistent as Opus 5.5 — 20:34

Concepts and entities

Gemini 4 Argon, Opus 5.5, Antigravity, Claude Code, Reverse Engineer Anything, Google, Jack Roberts, Harness

Full transcript (original language)

00:00 Google just dropped its most powerful model, Gemini Argon 4, and it beats Opus 5.5 on over 70% of benchmarks. So, in this video, I tested this Google model against Opus 5.5 across seven levels of difficulty for building the exact same website, app, and viral use cases like reverse engineering, so you can see exactly which is the best model to use right now. I’m going to show you what it cost,

00:30 how long it took, and the results were not close. And if you’re new here, I’m Jack. I built and sold my last AI startup with like a gazillion customers. Now I’m building my own AI startup, and on this channel, I share with you the stuff that actually works. So, if you haven’t already, grab that beautiful coffee, and let’s dive straight in. Now, having Google and world’s best model was not on my bucket list for 2026, but this one is a big deal. It is their best model today, and it is crushing it on many benchmarks. In fact, 12 out of 18

01:01 of the general benchmarks, it is number one. But guys, we don’t care about benchmarks. We care about actual results. So, in this video, I’m going to show you the tokens it uses, how you can access it, and what actually the results typically look like. There’s so many people are building with Argon, you can actually access it right now in the Gemini app. You can download Antigravity, which is Google’s own integrated development environment, looks just like Claude, or you can just use an API. So, let’s begin with level one, which is the task of creating a beautiful website. Again, they’ve got the exact same prompt, so let’s see how

01:31 they change. Beginning with the prompt, which is build me something for electric perfect. So, let’s go ahead and check out what Gemini did, and then we can see exactly what Claude coded as well. And bear in mind, we’re using all the image and video generation, we’re using Hexafold. I just gave them an API, and from that, they can just generate everything. So, this is Gemini. It’s got a basically like a kind of like a background thing. I can see a nice video that it’s generated that gives it a little bit of depth. Definitely has initially a little bit, I’d say, of an AI feel to it. I think it’s done a fantastic job of

02:01 generating good images. I think the text itself does present slightly less kind of AI sloppy. Come down. I’ve got a Codex thing here. That’s cool. We’ve got the perfume being displayed down there. Yep, okay. Come down, you can see it. And it’s got that kind of gold classic Gemini touch. But again, one prompt, this is exactly what it went and demonstrated. Then here, for example, we have Claude. Now, this is fantastic. Now, obviously, one of the cool things it did, it generated this beautiful video in the background. I mean, look at this. It’s like the waves are crashing upon the rocks. And by the way, all

02:31 these challenges are completely different. So, you have to stick around to see how it performs in all seven. Trust me. Now, what’s cool is I like that not only do you have the waves in the background, we’ve also got the smoke coming on. And it’s called Scent of a Virgin. I don’t know what that means. We’ll find that one out later. A perfume should not enter the room before you do. Cool. And we’ve got even got a lady holding it. So, we have product placement that’s going across both. This is a distance. Come closer. I can scroll down. And once again, we’ve got a beautiful video really bringing out the perfume as the highlight of the whole asset. Guys, you know what? I might actually end up

03:01 buying Scent of a Virgin. I think it’s a really interesting present. And we’ve got one here. Um after last delivery, different different sizes. Very cool. Scroll down. Yeah, very very interesting. At the bottom, we’ve got something a little bit unique. Now, I gave it in the references, I just gave it one prompt. And so, Claude did this in around 30 minutes. It took Gemini an additional 18 minutes. But the cost of Gemini was significantly cheaper pound pound in terms of its token use. So, of course, we have to go unbalanced performance, pick Claude for that one. Now, level two is really interesting.

03:32 Level two is a 15-second animated product intro. Something that Opus 5.5 is incredible at. Let’s see how well Gemini did against that. So, the prompt for this is essentially a 15-second intro for a habit tracker. We’re going to check out Gemini. We’ll check out Claude. So, let’s go ahead and see what we’re dealing with. So, this is what Argon developed. Now, animated graphics, it’s pulsing, it’s got the screens coming on. We even have the tracker. It’s got this beautiful multi-colored luminescence in the back. A big live story of his

04:04 ritual. Honestly, I’d say this is really impressive, to be fair with you. Like, this is a good crack. Obviously, we’ve got the screens coming in and there’s a couple of one or two things I’d change. I think this looks a little AI ask to me and the the bottom of the G is missing off, but I think generally speaking, this is pretty interesting. Not not a bad go. I mean, you can come down here and click through everything. Let’s check out what Opus 5.5 did. Now, wow. I mean, this is in terms of an animated intro, fantastic. Like, this is genuinely

04:34 ridiculous. Wow. And I think what’s cool is obviously I connected this also to Hexafold, so it generated the images in it as well, which is fantastic. But, even the way that it kind of actually created beautiful logo animation, which I highly recommend you do, by the way. I think this is clearly a cut above and a level above what we’re getting with Argon. Now, it’s said that Argon actually beats Claude on everything except the heavy coding. And if you put this in the heavy coding category, that would make sense. So, Argon completed that a little bit faster and the price of this, guys, was

05:05 significantly cheaper with Gemini, but I think from an actual performance point of view, we have to go with Claude as a clear winner. Now, a little little hack for you, if you are using Argon, I honestly found that you want to be doing it inside the anti-gravity app and you need to come down here and you can see you’ve got Gemini 4 Argon, you can select it. If you use Gemini 4 with Open Router, okay, it doesn’t use the anti-gravity harness and I can tell you right now, I found the results were so bad. Some of these challenges I initially did with the API, they just

05:36 weren’t good enough. So, I actually re-did them using the Google harness, so you can get a great idea of it. And that isn’t just Gemini, that also applies for Claude if you’re not using a good harness. Now, level three is interesting. And by the way, if you’re following along wondering how I’m doing this, I just gave it a Hix old API, which if you’re not familiar with, I’ll put a link down below for this. Essentially, what it does, it lets you access different models. All that I’m doing is clicking on API, and I just said, “Hey, use this API key, and basically just build for me any images or videos that you need.” And to follow this, just come over left hand side to API keys, and then just come down,

06:06 create an API key, and you can even call this one Gemini or test, whatever it is, come down and create the key. And then you can literally copy this, come over to Claude or Gemini, and then it will literally be able to access all the models. Or if you prefer, you can literally come over to this interface here, and just enter anything that you want to. But the reason why I’m using the API in these tests with Gemini and Opus, number one is cuz we get many models. So, whether it’s seed dance 2.5, um cling, whatever the thing is, we can get them all. Secondly, agents can use it at any point during the build, which is cool, so we can check the status of how that’s going, which is really handy.

06:37 And obviously, we’re going to get real assets for a completely fair task, cuz every model has completely different capabilities. Now, as you can see, this is the stuff that’s built, but the level three guys is really interesting, because we’re going to build now a Minecraft style game completely from scratch, 3D physics, controls, all in one file. And by the way, I’m going to tell you how I also built these. And if I double click, let’s go over to talk to Gemini now. Now, let’s see exactly what the prompt was. The prompt basically a playable voxel world in the browser. You walk in first person, or break blocks, uh world day turns to night. So, let’s see what Gemini 4 Argon did first of

07:09 all, and then we’ll check out exactly what Claude did. So, this is Gemini. We’re going to enter the world right now, guys. And as you can see, believe it or not, what we’re looking at here is I think Minecraft. I’m going to be honest right now, guys. I do not play a lot of Minecraft. But as you can see, I’m doing some interesting things here. So, I’m going to change blocks. I don’t know what the hell’s going on here, to be honest with you, guys. I should really play a little bit of my uh Minecraft the homework here. But you can see, it’s really good. You can play around. I can place blocks down. I would honestly say, this is pretty decent. I mean, Gemini did this whole thing. I thought this was fantastic, honestly.

07:40 Really good. I’m not rigged up for a gaming PC, but you can hopefully see the kind of things that you can do in this Gemini world. Very, very cool. That took 35 minutes and cost estimated $5. Let’s see what Claude did. Again, very, very comparable, actually. Let’s check out Voxel Island for Claude. And then here we go, guys. So, this is, if I show you right here, what we’re dealing with Claude. So, I think fairly similar. I mean, its ability to honestly build these worlds is is very, very impressive. I’m going to come down. I can click things. It’s managed to understand what the hell’s going on. It’s full of debris. Pretty decent. Honestly, I think this is fairly

08:12 equal. I think the world that it’s built in Claude probably a little bit more expensive, to be honest with you. But really, if you compare this thing with the other one we had at Voxel Craft, just so we can refresh our memory, I’d say honestly, I think probably the color grading’s a little bit better in Claude. But ultimately, you couldn’t really There’s not that much difference between them. And so, I’d say this was neck and neck. I’d probably give the edge to Claude just realistically. If you zoom in, I think Claude just did a little bit of a better job. So, let’s get our Claude guy out. And I think Claude is the winner there. Although the two It’s credited Argon did that at a

08:43 much cheaper rate. This takes us on to level four, which is a mini Photoshop free in the browser. Now, this is using, guys, a brand new GitHub repo. And this is called reverse engineer anything. It’s completely blown up all of the action on YouTube recently. It’s now got almost 300,000 stars. And effectively, the idea is that you can give it anything you want to and it can reverse engineer it. So, it even reverse engineered the entire Adobe suite. So, what we did is gave both of them that access to see what they’d come back with. So, the prompt was basically, “Go ahead and build me a photo editor in the browser.” So, let’s check out what both

09:14 of them did and see how they performed. So, this right here is Luminar Pro. This was developed by Gemini. So, let’s take a look at this. Honestly, and I’m going to tell you right now, this was one prompt. I didn’t do anything at all. I can change the uh various different like filters here. I can play around with the brightness, which is cool. I can play around with the contrast. Let’s give it an an image of me, for example. So, it’s going to come into the file and perhaps upload open image. And I’m going to upload an image that I I use for Claude. So, this is our beautiful bird we’re talking about, hey, you know, we’re fast at dictation cuz we have command mode that can do really cool stuff. And let’s say I want to amend the image. Do I

09:45 brightness? I do, very cool. I can play with the saturation, that’s cool. All these different elements and things I’ve left on the side. I have filters, I can rotate it, crop it, do whatever I like. This is a very Photoshop-esque thing. Cool. So, I think this is fairly good. I can control zoom way out of it. It’s kind of like a basic functional Photoshop. I’ve got the brush, so I can kind of like draw maybe a little circle into it. Very interesting. Obviously, and they’ve got the different layer section down there in the bottom right. I would say that it’s a little cluttered, but honestly, for a one shot, I think it’s really good. And you could probably if I just explained what I

10:16 wanted, it would do really well. Let’s check out what Opus 5.5 did. Now, Opus 5.5 initially, I think just kind of starts a little with a little bit of overview of the user experience. Now, let’s go ahead and try an example here by giving it the same image. Cool. So, we’ve got the Claude image here, which is cool. Now, it’s got a slightly less functionality, I think, in terms of what the adjustments are. We’ve got filters. The filters look a little bit better. We can play around with that, which is fantastic. The layers are good. Can I add a layer? I can. I’m saying, “Hey.” All right, so we can I can move that around potentially. No, it

10:47 doesn’t seem like I can move that around as easily. I think I need to come up to this button here. Yeah, and it is cool. And if I just drop this above, it’s worked just like Photoshop, and I think it should disappear. Well, let me bring that over. No, it won’t because it’s locked. Man, I will tell you right now, they’ve done a really good job of kind of emulating what Photoshop would be like. Bit of a simpler interface. I think the UI is a little bit cleaner, it’s a little bit crisper. So, I would say from a UI perspective, I would give that to Claude. I thought Gemini for Argon did have a little bit more functionality, but overall in user experience, I’d probably say Claude. But

11:17 to be honest, guys, if I just now look at the cost, which is one of the things I really want you to take away from this video, is that performance is one thing, price is something else as well. And there’s a few other things that may surprise you in this as well. But look, it did take a little long to do, but the output per token was significantly better with Gemini Argon compared to what we got with Claude code. That said, I think the final product edge I probably would say is 50/50. But I’d say I’d have to again give this one to Claude just for kind of making that UI experience a bit better. And remember sometimes guys, I would say that less

11:49 can be a lot more. Which is why we’re looking now at level five, which essentially is something personal, which is building an app built around my week, has to understand my life. So I gave it a brief like, “Hey, here’s the TLDR on Jack. Go ahead and basically build something that’ll be helpful for me to use.” So really test it. And I think like oftentimes a good way to test performance and capability is how little information can you give a model, right, and it’s still perform and do its job. So how little context do you need to do to do a perform excellently. So the prompt, go ahead and do something about

12:19 Jack, and let’s see how they both performed. So we’ve got this one, and then we also have Claude. Okay, so this is the green light operating system, and sweet moogly googly, this is outrageous, guys. Oh my gosh, dude. Have you ever had like hierarchy like you know one page one thing? I have absolutely everything fighting for my attention with its life on this page, guys. I don’t know where to look. It’s like someone said, “What feature do you want?” And they just simply said, “Yes.” This is unbelievably complicated. I’ll be completely frank

12:50 with you. This one is outrageous. It’s not really done the thinking. And I this is what I’m realizing with this stuff like you know I deliberately kept this one vague. This is pretty crazy pretty outrageous. And the thing is I think there could be a lot of really good intelligence going on here, but the issue I’m getting unless you’re so specific and I think get a little more AI slop text here, it is just going so far around the houses with this sort of stuff. It is crazy. But I mean look at the detail. I mean it’s got outline packaging, deck and asset studio, sponsor approvals, edit thumb queue. So I think like functionally like it’s

13:21 thinking about it in the right way, but the way it’s being presented is just out out control. Now let’s compare that with what we got from Claude. A little bit easier to work our way around. We got this week, so organizing everything this week, which I think is handy. Still a little busy, if I’m completely honest with you. I can move things around. It’s kind of built a really nice thing I can move around. Honestly, it is doing a good job. We’ve even got like some images generated there, too. We’ve got ideas. We’ve got sponsors coming down. It’s done a really good job, to be fair. And the results, I think just genuinely easier to understand all the way through. So, I think this is a pretty

13:51 uncontroversial win. So, clearly, this one here also has to go over to Claude, guys. And again, I think it’s cool it can just go ahead and design the stuff. Now, this takes us on to level six, which essentially is to ship a feature in my own dashboard. Let me show you what I mean. Hey there, Glido. Open up my Agentyc Operating System Jarvis. Now, the cool thing about this Agentyc Operating System is I can manage everything from Claude and ChatGPT and CodeX. I can look at all my different agents, right? So, I’ve got a Grok bot, a Hermes bot. Everything’s happening here, and I have my full memory system. So, one of the things that I wanted to

14:22 do, cuz obviously, I’ve got everything here that I’ve saved. I can chat with everything. It is incredible. Saved me so much time. But, what I wanted to do is take a look at this and say, “Cool. Well, what could we potentially do? Like, you know, if I give you carte blanche, like go and find something interesting, unique, just to test, like even outside of the creative output, what can Gemini and Opus 5.5 actually do if I say, “Go and find me something interesting.” Now, I’m going to come down and show you, okay? So, we’ve got the prompt here, which is another dashboard feature that helps me every day and matches the style. Let’s see what they both delivered for us. So,

14:53 guys, this one here actually is going to be from Gemini. Now, I’ve just got a Jarvis here to help us guide it. Hey Jarvis, we’ve done some interesting developments here to our production state slate and outcomes. Could you give me in just a sentence or two, what is new here? So, what am I looking at here with videos production waiting on thumbnails? >> You know that show? You’ve got a lineup of videos produced and ready, but they’re on hold because the thumbnails aren’t done yet. It’s that classic waiting for the final polish. >> Got it. Thank you, brave lad. He’s helpful for talking about things. So,

15:23 what I’d say I’d say design cluttered a little bit, I’d say. What it’s tried to do is help me streamline my view generation and pretty interesting. Let’s check out what Opus 5.5 did. I come over here. Same kind of thing, but I think again it’s visual design skills have just kind of edged it out here again. In the making and published, I can kind of see where I’m at with everything. Have I recorded it? Where am I with the deck? I can see this being really helpful and this is one of the cool things about Agent OS, guys. Whether you’re doing websites or designs, whatever it is, you can really build it

15:53 and mold it around whatever the thing is that you’re actually trying to build. So, I think that the answer there very clearly in terms of the winner has to go to Claude. Now, let’s look at the top the cost and the time. You know, Gemini took longer, Argon took longer, but from a cost point of view, guys, wasn’t even close. It did it in a third of the tokens. So, what I’m getting at about this is all one prompt. So, I think if you gave it another prompt, a follow-up prompt, I think you could approximate this design very easily. But, it takes us very nicely onto level seven. Now, level seven is a very bold creative one,

16:25 which is designed this thumbnail for this video. And guys, every image and video that we’ve generated from the websites to even actually the assets in these games like Minecraft, we’ve just done using the Hexel API, which is why it’s so handy to have one API key or CLI connection and build these. Now, this one’s a really interesting one because it requires an understanding of what makes a good YouTube video and what makes a good YouTube thumbnail. So, see how it does it a completely creative test. So, right here we have Gemini and sweet higgledy-piggledy. This is crazy. Now, this I think is not great. But,

16:58 never mind. This is not good at all. This is really crazy. I mean, I think if I look at what it’s done, it’s kind of had the right idea. I think actually is this based on a viral outline? If so, I am really impressed with that. But, I’m not like it’s kind of like accomplished what it wanted to do, which is cool. So, it’s it’s kind of built a little bit of a system around it, which is interesting. On the top of this, also there’s a a teaser trailer, which is pretty incredible. So, have a look at this. You can see it’s got a lot of stuff happening. Yeah, I mean, pretty

17:28 not incredible work, if I’m honest, from Gemini. Let’s see how we got on now with Opus. Now, this one then if I come down to have a look at Opus, let’s see what it’s done. Now, Opus itself hasn’t had a great idea of what makes a good thumbnail, but as you can see, I’ve come down here and check out some different ideas. It’s got a split. It’s describing things. It’s not really that fantastic. Let’s see how it did on the animated trailer. I’m going to come out and just show you this one. Let’s come down. Seven levels. Okay, let’s have a look at this. And just explodes with two balls attacking each other. Only one makes it.

17:59 It’s kind of done like a cinematic trailer, rather than something it that you could look at in particular. So, you know, a complete stab. Just goes to show that you can pretty much generate anything with this sort of stuff. Neither of them blew me away. But, I think in general, if you come back and look at the prompt, effectively what I said to them here is just research what makes AI versus thumbnails good and then build me something. And I think it just goes to show that sometimes you need good knowledge to make good prompts. Well, before we go into the final stage with this, if this all sounds like I’ve been speaking Spanish, or you want to turn X, I’ve got a link down below for

18:30 my full Agentyc masterclass that will take you from a complete beginner through to learning how to build beautiful websites and design systems and take things to a completely new level. And you also get immediate access to the entire Agentyc operating system or memory systems and all this stuff uh that you can unlock straight away. Now, one thing I really want to double click on here is the fact that Claude itself won seven out of seven of those results. Like, absolutely won seven out of seven of those results. With one being a bit questionable. But, the real question is

19:00 what is it worth in terms of economics? And I wanted to answer this question because I was playing background anti-gravity. I actually really enjoyed, to be honest with you, playing around with anti-gravity. Like, it was kind of fun to get back into it. The interface is really nice and crisp, and I was really rooting for Argon 4 to win. I really wanted it to surprise me. Now, when you look at the face of comparing Claude to Gemini, one of the things that you see is that Gemini itself, okay, will get you more because it it is a little bit more token efficient, but that token efficiency is crushed when you actually have to go back and

19:31 re-amend it and congeal it and move it in a thousand different directions. So, this gives you an idea here just how far you can expect those subscriptions to actually go, whether you got the AI Ultra $100 plan or the Google AI Ultra $200 plan. So, for literally double the price, you actually get 4x the usage, and they do have basically resets as well. And just to give you an idea, guys, just building these apps today only took about 25% of my weekly usage with this as well. And that’s on the $100 plan, just to

20:02 give you an idea. And that’s classically been one of the things that hold back the anti-gravity in Google’s system is what they did with the usage. That said, it is more token efficient, but you’re just going to have to do more with it. But, if you look at everything, guys, we spent 4 million tokens on Argon, and Claude spent 22 million tokens to get this thing done. But, to access Argon, you’re going to need to actually have Ultra because they haven’t included it in Pro. So, with all that in mind, Jack, what is your big recommendation to users? Honestly, guys, there is no reason right now to leave Opus 5.5 or

20:34 ChatGPT or OpenAI right now where it’s at. What this has shown me is that the benchmarks, which we’ve always known, don’t really mean that much. Like, yes, for sure, you could use these to do certain knowledge-based tasks. Of course, you can, and it’s probably a good upper middle pack model that will just work nice nicely for you in the background. That’s my take on it, but it’s not quite at the frontier edge where Opus 5.5 lies. I think that the difference in some of these was quite stark, but yet in some of them it did quite similar. It’s just not quite as

21:06 consistent as Opus 5.5, but it does bring us on to a really important question. And that is what happens after the model. After the model, we need to know exactly what to build today. What is the stuff that’s actually working, which we’re going to cover together in this video right here.