Does GPT 5.4 beat Claude Opus 4.6?

Hey there, my gorgeous friends on the internet. For the past couple of days, I've been running a bunch of experiments to see the strengths of Opus 4.6 and GPT 5.4 on high, and I have some thoughts to share with you. So, let's get straight into it. So, the first prompt I tried out uh was creating this SVG plant animation. Uh, and the prompt I used was build a single HTML CSSJS demo using SVG to simulate a plant growing stems extended leaf sprout and do some wind physics as well to it. And here are the results. So, I feel like Opus tried way more to focus on the physics aspect and then it just ended up glitching out unfortunately. So, as you can see, you get like the jaggy effect whilst GPT 5.4 four really got that like leaf physics effect. So, see how it pops out? That looks really satisfying, but I kind of messed up the positioning on some of them. So, as you can see, you have a leaf on top of another leaf. Why I also specified in the prompt was that it should loop, right? So, Opus 4.6 here just decided to do a fade out. Uh, whereas GPT 5.4 has like this kind of plan dying effect and then coming back to life. So, at least it tried it. So initially I was really excited because I was like okay GPT uh has you know has a good leg to stand on. Is that a saying? I'm not sure. Uh cuz I primarily used Opus 4.6 so far uh for my daily work. So good test. So I gave this one to GPT 5.4. Now I also ran all of these tests about a month ago on Gemini 3.1 Pro as well comparing Opus. And I'll just quickly show you this uh little example here. Uh as you can see, Opus really like tried to get this physics right. Whereas Gemini 3.1 Pro looked prettier. And that that was kind of my experience with Gemini altogether is that it tried to make things look pretty, but often times uh the code wasn't properly working. It would run into a bunch of syntax errors. It would have real hard trouble uh with tool calling as well. So, I haven't really used Gemini 3.1 Pro, but anyways, let's move on to the second test. So, for the second test, I tried to prompt it to essentially make me a nice landing page for a service that I can sell. And here are the results. So, this is when I started noticing right off the bat with the first test is that GPT 5.4 does a great job. It like codes really well. uh if you want to implement a feature a lot of times you know I have a look at it I'm like that's better than a lot of developers would actually do. Uh so really impressed with the coding but I noticed that when it comes to UI it tries to overexlain as much as possible and it adds a bunch of text adds a bunch of cards and ends up making things look really busy. So as you can see Opus 4.6 six here has a more clean and minimal look to it, bit more modern. GPT 5.4 feels a little bit like it's one gener generation behind in terms of UI design. So here as well, see how minimal it is even on hovers, whereas GPD 5.4 just does these crazy glow effects on it. So there we go. Uh I'll let this play out a little bit so you again see like how crazy is that glow. It just doesn't look good at all. Other than that, it was really, really close. So, that was the second test. Moving on to the third one. I wanted to see how they do with 3D models. One of the tests I like to do is to create like an isomorphic scene, uh, where you can click on different items and you get a nice little description about it. And I'm not really sure what happened here with Opus, but it just completely failed. Now, if we play this, as you can see, it made the scene quite dark as well, whereas GPD 5.4 for is a tad bit more appealing. All of these like 3D assets were made by both models as well. And Opus 4.6 here just just didn't manage to get the camera working and it glitched out and would just never stay in frame. Um that's even after like two reprompts. It still had a bit of tru uh trouble with it. But really liked how it made the the little nav bar here as well with all the different items that you can click on. So, I don't know, maybe Opus was having a bad day, but GPD 5.4 absolutely smashed it. Now, I looked at the actual assets I created. And if you just pause here for a second, even though this had all the glitchy effects, I feel like Opus actually did a better job with modeling these. That is supposed to be a bed there on GPD side. Now, I don't know about you, but that looks a bit weird. Now, here, that looks like a proper bed, right? You got the pillows, you got the cover, everything looks great. Look at the the little bookstand here as well and the laptop and the cacti. Really nice. If you look at the cacti here on this side, it's quite blocky and it's kind of positioned randomly kind of like the leaves in the first demo. See, that looks kind of crap. And the laptop here looks kind of crap, too. But overall, it did manage to pull out a better scene. So, it was two for GPT, zero for Opus so far. But don't worry, that's all about to change. For the third test, sorry, I guess it's one for Opus cuz uh this was way better. Now, for the fir fourth test, I tried to do a 3D flight simulation where you have a planet and then you have all the flights that are getting pulled from an API and you can track different flights, stuff like that. So, this is the result. What's was really interesting is that initially GPD 5.4 did not do a texture whereas Opus had big absolute big brain moves and went out of its way and found a GitHub repo that had high quality textures of the planet and just put it on there. So I was blown away. That was really really impressive. Other than that the texture, it might not look great on the screen uh but the texture here looks way way nicer. This looks quite faded out. Um, and the actual like um airplane models, as you can see here, GPT is also like clipping through the planet and it just didn't really get the models properly. Whereas Opus again like got the models looking pretty good. That looks like a plane to me and it even has the lights on. So, Opus one again. Um, I noticed that when it actually comes to making more 3D complex models, it just kind of falls apart a little bit. Opus actually tries properly. For the fifth test, this was again I tried another UI test to see how it would handle something a bit more complex like a stock analysis app because you have dashboards there, you have time simulation, you have indicators, graphs that you need to render. And this was the result. And again, I noticed that Opus just looked far more clean and the actual charting system here was fully functional and it was really smooth as well. Whereas GPT I'm not sure what it tried to do. Uh but overall even here like GPT just tried to cram as much text in as possible and just didn't even make this properly uh mobile responsive. Uh, but you're going to get down here to this section here. And as you can see, we have overview, financials, valuations, technicals, blah blah blah. Uh, look at how nice Opus looks compares compared to GPT. So that's the technicals, you got valuation, you got financials, and you got overview as well. GPT just again just pop everything in a card with some text on it and call it a day. That's kind of how it does. So I actually much preferred Opus in this one. For the sixth test, I wanted to see how they do in cloth simulation. So I determined that Opus is the winner when it comes to actually modeling stuff. And that's pretty clear by even like this plant versus this lamp here. Uh, but when it cames to actually getting the physics right, GPD 5.41. Not only that, GPD 5.4 optimized the scene properly to make this buttery smooth, whilst Opus I I felt like it had some like frame rate issues. So, I was really impressed. I also told it to be able to cut the fabric. As you can see in GPT, it just like works perfectly smoothly and you can grab the piece and throw it around and everything. whereas on Opus 4.6, it didn't really work and it just kind of stayed floating there. Um, so not sure if Opus tried to do some more more like post-processing on it that just made it quite laggy and janky, but GPD felt really really nice. Um, and then I wanted to see how it does if we reference an actual image. How how well can it replicate a UI? So, I found this little image here that had a light mode and a dark mode. And by this time, I I kind of had a feeling that I knew who was going to win by by this test, and that was going to be Opus. So, that's the reference image here. And if we have a look at it, they're both pretty close. Not really a big fan of how GPT did the the clouds here. This looks much more appealing in my opinion. And Opus was also smart enough to toggle to just like use this one circle shape and kind of morph it into the moon. Whereas I noticed GPT 5.4 just kind of ignored it and didn't really consider it and just ended up making two circles that it just faded in and faded out by. So much much prefer this. Much prefer the stars on it. The way the clouds come in here is is a bit janky on GPT 5.4. Overall, I I kind of knew that Opus was the way to go. And then finally, I also ran a test to see how well these LLMs help visual learning, right? So, if I want to do something like, okay, I want to explain bubble sort to you or quick sort, which model should I use? So, same same situation. That was the prompt. And look how minimalistic Opus has done it, right? It knew the goal, which was to make some sort of visualizer. And you have a little bar here just at the bottom with play, pause, uh, to make the animation faster, scrubbing, or add new data points. Look what GPT 5.4 does. It adds as much text as possible. And there was way more text here at the bottom as well with as many cards as possible as well. it's just not too too appealing. And when it came to actually doing the animation as well, I it was it was a clear win here for Opus again. So just my final thoughts on this is and using it for my project as well. Oh, speaking of ask if ASI genen we released version 0.5. Now, what that means is that I added a bunch of adapters. So, you can export your ASI animations in React Zelt as well and vanilla JavaScript. And we have color mode as well. So, you can literally go upload a video, an image or a image sequence. Let's just I'll upload a video. I have a video here of a fish to show you. Uh let me go to my downloads. Here we go. We got a fish. All right. We'll load the fish in. And as you can see, that's a little preview there in ASKI animation. But if we go down here, we can add color. Look at that. We have a full colored fish. How cool is that? And then we can export it and whatever. And then add it to your React app. So I was working a lot on this with both Opus and GPT 5.4 as well. And GPT gets the logic right. Okay. You don't need to worry about it. And I actually highly recommend that if you want to use use both of them. Get both subscriptions and just let Opus do all your design decisions and implement that first and then you can switch over to GPD 5.4 uh for the actual coding part of it. Now, one more thing I want to say is the actual usage you get out of codeex compared to um clawed code here is ridiculous. So, both on the $20 plan, I noticed that you can get maybe 5x more out of codeex than claude. Sometimes I with one of these prompts here. Let me show you uh because I also added the usage. So, for this prompt here to create this landing page, Opus used 37% while GPT 5.4 on the codeex plan used 3%. That's insane. And they're both $20. time was around the same in all the tests. So, I wouldn't really worry about that. But don't just do everything on one model because they have their own strengths. So, yeah, that's about it. Um, I'll be doing a bunch more. I'm just doing some tests now if fast is worth GPT 5.4 fast. So, come on X and uh I'll let you know. Okay, that's going to be it for me today. Hope you enjoyed this little episode and check out ASKI gen art. It's really cool.