Claude Opus 5 vs GPT‑5.6 Sol: tres proves, un guanyador i molts matisos

So, Claude Opus 5 launched earlier this week, and what can I say? It's a good model. I compared it to Fable recently, and it won by a landslide. However, there's one model in Anthropic's own benchmarks that Opus 5 couldn't beat, GPT 5.6 Soul. Every benchmark I've seen online ranks these two models extremely closely. So, today we're putting them to the test by giving the exact same prompt to both Opus 5 and GPT 5.6 Soul across three massive builds, a motion graphics video, a fully playable first-person shooting game, and finally, an interior design app. Same rules for both, a single one-shot output, no revisions, deployed live to the internet. This is Claude Opus 5 versus GPT 5.6 Soul. Let's get straight into it. Up here on the left-hand side, we have GPT 5.6 Soul on extra high effort. On the right-hand side, we have Opus 5 on extra high effort as well. Now, I recently hired an animator to create intros for my short form. Check this out. >> [singing and music] >> So, what I want to do is I want to see if I can get Claude and Codex to create something similar for my YouTube channel. And we're going to be doing this using HyperFrames. Because using HyperFrames, you can actually get your AI agents to create motion graphics using code. And what's really cool about HyperFrames is it's completely open source. So, you can literally go and try this today if you have an AI subscription. So, by creating motion graphic intro animations for my YouTube channel, this one build is actually going to test a lot at once. We're testing design taste, typography, animation timing, and also frame-perfect code. Because this can't just be a text slide across the screen, the models have to write the whole animation in code, every [snorts] letter, every shape, every transition, timed to the exact frame. So, this could actually be useful if you're making your own intros, running a faceless YouTube channel, or paying an animator every month like I am. Let's see what these two models cook up. I've got the prompt right over here. Use a HyperFrame skill to create a 20-second animated intro for my YouTube channel called Dooby Dooby. So, I'm just going to go ahead and copy this. And if you want access to all the prompts that I'm using in this video, I'll leave a link to this document in the description. Boom, both models are now running. And of course, you know the drill. We're going to be using this usage dashboard to verify everything. We're going to look at the in tokens, output tokens, the time it takes, how much it's going to cost us. And guess what? This dashboard is now open source. So, if you want to run your own test, then I'll leave a link in the description, or you can just put this into your tab to use it. It's really simple. Just paste it into your AI model and just say, "Hey, I want to use this dashboard. Can you set it up for Claude versus Grok?" Or whatever models you want to use. Now, for the scoring rubric, I have Sam as a wizard. He's representing GPT 3.5 Turbo, and I have Boris Cherny, the boxer, representing Claude Opus 5. Now, last time we did this, you guys seemed to really like me doing the blind testing. So, I'm going to do it again, and I'm not going to know what models generated which outputs until after I've tested both of them. Because it's going to show as a little thumbnail in our dashboard. Anyway, I'm going to let these two models cook. I'm going to go for a walk, and I'll see you in a bit. Yeah, I'm not even going to lie. I'm extremely curious as to who's going to win this bench. Because in my own personal use, I'm constantly switching between Codex and Claude, and I find them both to have strengths and weaknesses. For example, Codex feels like a scientist in a lab. He's very analytical, straight to the point, and just gets [ __ ] done. Its ability to execute is amazing, but it does tend to falter at times. Claude on the other hand is like a really hot girlfriend

that won't shut up. It left a really good first impression on me. It one shot at a bug I hadn't been able to get fable to fix after weeks of back and forth. However, it's extremely chatty. I find myself reading essays after Opus completes a tiny task. This makes skills like I have ADHD or the caveman skill increasingly more useful as they teach your models to cut the [ __ ] and give it to you straight. Anyway, the models are probably done now. I'm going to finish this snack and then we'll go back and have a look. Imagine how far back you got to walk to collect that camera man. All right, both models are finished. Let's see how they went. I've got them both lined up here. Let's start with this first one. Woah. Woah. Ooh. That is really cool. Woah. Damn. That is a sick intro. That was really freaking cool. I'm not even going to lie. That was dope. I really like how everything was like to the beat. And look how much animation there is. Like look how they come into frame. Signal. Like this is sick. Look at that. That's such a cool animation right there. Boom. How it explodes. Super cool. And then you can kind of see it's like building up to say the full channel name, which is Dooby Booby. That's actually going to be really hard to beat. That might be one of the better generations that I've seen from an AI agent. Let's have a look at number two. >> [music] >> Cool B. Full power. Uh Yeah, I don't know about that one, man. I can almost guarantee that this is GPT, and I'll tell you why. GPT has image generation capabilities. So, you can see here it's generated an image, and then it's tried to use that image in the video, but it just looks kind of lame. All it does is zoom in a little bit. And then you can see over here it's done that again with this weird texture up here. I don't know, I'm not really vibing this one. Let's have a look at these stats. We got Damn, that's a big difference in token usage. 39 million and 15 million. And if we look at the thumbnail, that looks like the first one right here where it goes just build Yeah, right there. Boom. So, that's Opus, that first one. The second one is GPT 5.6. And that is a big difference, man. 25 minutes, an hour and a half. $28, $11.42. And you know what's crazy is I totally expected that because if you look at the benchmarks, the cost difference between GPT 5.6 Soul Max and Opus Max is literally is quite a big jump. I think it gets even worse when you take in the fact that it's way more token hungry. So, you can see here it's nearly done three times as many tokens compared to GPT, therefore costing nearly three times as much. Now, GPT's wasn't bad, but I still think Opus's was marginally better. Not only was the music actually catchy, but everything was on beat. I was actually really impressed. So, I'm going to have to give this one to Opus, man. I like you, Sam. I'm going to have to give you a little frown in face. Looks like a mustache. Anyway, let's go to the next prompt. Voxo first-person shooter. I'm actually really excited for this one. Create a complete fully playable first-person shooter in single HTML. I'm going to go ahead and send this prompt

to both agents. And I'll see you when they're done. All right, both models have just finished the build. And you can see over here, GPT only took 10 minutes, which is actually kind of ridiculous. Let's have a look at the build. All right, Voxo blaster. Which one should we do first? Eeny, meeny. Oh, this one's quite colorful. Maybe we start with this one. WASD to move. All right, let's do it. Whoa. Oh, [ __ ] We got monsters. Oh, oh, damn. Oh, cannon. All right, this is pretty fun. Oh, [ __ ] And then you just got to try and survive as long as possible, hey? Is this like a like a patch? Oh, so that must be like ammo or something. Oh, this is a laser, sniper, whatever you want to call it. Oh my god, I've found the hack. Boom. Boom. For a one-shot game, I'm having a lot of fun, guys. And goes to show, anyone can make a game like this. You got cubed. The arena claimed another. 81 kills, six waves. I mean, that was fun, man. That was actually really fun. All right, let's let's give the second one a go. Voxo blaster. Survive the waves, clear the arena, break everything. Oh, okay. This one's got music. Oh, someone's behind me. Let's go up there. Oh, they can come up here. Oh, no. Oh, [ __ ] He went over, dude. That's kind of scary. Imagine I just get really good at this game, and I'm just like the best at this game in the world cuz no one else plays it but me. Oh, there's a boss. Damn. Oh my god. Okay, I healed a little bit. Oh, there's so many monsters. [ __ ] Oh, I died. Oh, that was fun. That was actually really fun. All right, let's have a look at the stats. Damn, Codex 2.7 million tokens. Okay, so it did that first one. It did that first one cuz it had the the brighter colors. 31 million tokens on that one. Opus 5 is super token hungry. It took four times as long and damn, Codex popped that out for like 1/10 the price. I'm sure you can see a trend coming on here, right? Where Codex is clearly faster and cheaper. It comes down to like what do you care about? Personally for me, the price difference is not that big of an issue if the output is better because time is money, right? And so if Opus is producing a better output that is saving me time, then I'm willing to pay a little extra. I know that people watching this are going to have a different opinion to me and they're going to think, "Ooh, you know, Codex is so much better than Opus because it's cheaper and faster and like you can just iterate and and provide a better output. But like, dude, for me, I'm just like, "Damn, it got the job done, you know?" So that's another point to Opus. >> A few moments later. >> For the time and for the price, I think we can't deny it. GPT did a pretty good job there. Let's get rid of your mustache and we're going to be doing things a little bit differently this time around because for the last build, you're going to be blind as well. You're going to be guessing with me which models generated which output. We have the interior design room planner. Create Room Craft, a complete fully functional 3D interior design room planner in a single HTML file. I'll send these prompts off and I'll see you in a bit. Okay, now remember you're doing this blind with me. So we're not going to know which models did what. We have both Room Crafts done here. Maybe let's start with this first one. Design a room you can walk into. Oh, this is kind of cool. Oh, the rotation is quite smooth. Wow. Oh, I can change the color of the walls. Halo side table. Oh, got to add a bit of that. Oh, and I can just put this anywhere. So, there's cool different types of views I can do here. Floor plan.

I think 3D view is the best though. So, we got different modes here. Industrial. Oh, that looks pretty cool. And it tells you the total room cost, which is kind of cool. On the left-hand side we have different We can get another olive tree or a ficus tree or I mean, they all kind of look the same. But industrial is kind of cool as well. Bit of a vibe. Leaning art. It's kind of cool. It's not really leaning. Japandi. It's feeling very Japandi. Anyone feeling Japandi right now? And then we got mid-century. Okay. I like it. They all kind of look the same a little bit, but still pretty cool. Let's have a look at the second one. Room craft. Design a real room in 3D. No manual work. Why Woah. Oh, is that like a skyscraper in the background? That's kind of cool to signal a window. We got a table here I can move around. I like this pulsing circle to show that I've selected it. Kind of feels like SimCity a little bit. Each one has a photo, which is really cool. Compared to this one where you're kind of guessing a little bit. Let's have a look at the plants. The plants look nice. Oh, a little succulent. Little succulent over here. Oh, put it on top of the table. See how it recognized that? I don't know if this first version had that. Let's go ahead. Let's be on the floor. It It has to be on the floor. So, this is pretty pretty cool logic right there. Oh, industrial is pretty clean. Let's have a look at Japandi. Oh, Japandi is a bit of a vibe too. And then we got mid-century. Ooh, mid-century is not for me, but I like it. So, if we were to compare this now to mid-century over here, I'm thinking this one looks way nicer. Can I put tables on top of tables? No. I can't. Okay. Well, what are we thinking, guys? I'm thinking this one's a winner. So, can GPT bring it back? Let's have a look at the stats. Okay, so we can see here that looks like the first one. Damn. Did you get it right? Did you guess that that was going to be GPT or did you think that was going to be Opus? You can see it's done 3.1 million tokens, 11 minutes, man. Look at the trend here, 25 minutes, 10 minutes, 11 minutes. Opus taking 43 minutes, 43 minutes well spent if you ask me. That apple looks way better. Opus total cost, $91. See here, that was all Opus. Again here, one nearly close to 1/8 or 1/7 the price of Opus, but the result is significantly better. So, if we look at the total use, I've only really used 12% of my current 5-hour window and 6% of my weekly, okay? So, are you willing to spend 12% of your total usage to get a significantly better result than GPT. But hey, I mean, GPT I've got 100% of my weekly usage left, so hey, I don't know. What do you care about, man? Are you on a budget? Do you care about iterating a couple more times so you can save a couple dollars, or do you just care that the result is good off the rip? If you're like me, then you probably care about the result being good, so that's Opus 3, man. Was this the result you guys were expecting? Let me know in the comments below. And if you want to see more of this style of video, then feel free to leave a like and subscribe because I make videos like this every single week.