GPT-5.4 Codex is Actually Great at Design

OpenAI's latest model brings native computer capabilities, allowing agents to operate computers, click through UIs, and handle tasks like emails, data entry, scoring 75% on OS world. We'll get into that. Matching human professionals across 44 occupations, command AI. You mentioned earlier on you're you're very confused about the naming. This might actually resolve that naming. So, GBT 5.4 Pro has also been released. So it's GPT 5.4 and GPT 5 um 4 Pro brings together the best of our recent advancements in coding and agentic workflows into single frontier model. It incorporates the industry-leading coding capabilities of GT5.3 codecs while improving how the model works across tools, software environments and professional task involving spreadsheets, presentations and documents. It incorporates the industry-leading capabilities of codec. incorporates. So that this has these coding capabilities built in. So really, if you're using 5.3 codeex, you might as well flip to four because it it is they've kind of baked it into one. 1 million uh tokens context. You know what? I haven't been keeping track of uh GBT on its token uh window, but 1 million is is I guess quite good. And it's our most token efficient reasoning model yet. And this is the thing about GPT, especially when you go X high, uh it it thinks, overinks, you know, it thinks for too long and then gets itself all worked up. Uh SweetBench Pro, which is the one um I'm interested in, this doesn't seem so high. So maybe Sweench Pro is a different um thing. Uh but you can see on its coding score, which is the thing I care about, is higher on uh than than GBT3 3 codec. So really cool how they focused a lot more on the technical capabilities of this tool. The interesting thing is it has the same knowledge date as GPT 5.3, which makes me think that they are using the same training data, but they're doing something with reinforced learning to actually kind of tweak the model and help it help improve it based on the same training data, which I guess might be which might explain a lot of these point releases that everything GPT5 point is all the same training data just slowly tweaked and manipulated and stuff like that and and you notice they're all doing it. So it might be that actually the uh training just costs so much frigin money that they're trying to avoid just training things up every year. We're sort of seeing a bit of a just training it and training and training isn't actually reaping the rewards. And I will definitely be using it in codeex which actually we were going to do a test, weren't we? Yeah, we want I was I was going to I was going to because here's the thing that people have been noticing about GBT 5.4 codecs. It's not great at front end design. And my argument here is that they're not really giving it the best uh opportunity for success. they are just telling it or or assessing its design capabilities by giving it a single prompt and telling it to build me a a dashboard or whatever. So, I've got GBT 5.4 selected here. Let's go let's go high. And I built a frame IO clone because I need to get my video signed off and all the rest of it. Um, I'm going to give it a skill, a front-end design skill, because I think this is really important that we we learn how to [clears throat] give the LLMs the task or at least the the abilities to to do that. Give me an exper experimental or do we go clean and cool? Do we do like modern and sleek? Yeah, but uh add some nice interactions and design elements. I don't think that's design though. I think we can do that, but I don't inherently think just cuz it's got interactions means it's design. We're f we're focusing on the design of the app. So that's color spacing font. Color spacing things like this. And and does it have taste basically? So um let's go tell it to be tasteful in some nice uh animations while you're at it. So, give me a modern, sleek, uh, modern, sleek, and tasteful redesign of this app. Boom. Should we let it Should we just let it do its thing? Cuz it might take a little while. Yeah. And come back. On high, interestingly enough, you I'm on high because I I I let it run on e uh extra high all the time pretty much. Yeah, I I I I normally do, but I don't think this is an extra high task to be honest. And um I don't want it just to take ages and ages and ages. So, let's uh it's got the scope. Let's let's come back to it and let's see where it's going because whilst we're on the codeex, we might as well talk about this. They have registered over a million active users, weekly active users um since launching, which is quite crazy. Um growing 400% since early 2026. I mean, I'm using it, but I do have You said you didn't see this, but I thought that they've been giving away double usage. There's like a a deal they've been giving out. The only thing I could think about is this tweet that I found from Samwin. I mean, it really doesn't matter too much, but uh we'll keep codeex available to free and go users after this promotion. We may have to reduce limits, but we want everyone to be a try try codeex and start building. I'm pretty sure I saw something that that um they're extending the the promotion or something like this. I don't know. But either way, I've been using it pretty pretty heavily and and you were saying that you've you've actually well, have you have you moved over to Codeex? Yeah, I'm using Codeex app for pretty much everything because of the [music] the rates of Yeah, the usage with cloud code. Um, I do like run clot code for one project uh at a time because like I can barely work through one project at a time before it running out. With codeex I cannot even get close to even like using 50%. Right now I'm at 72% of my weekly and that's after like tons and tons of use. um all on extra high many tasks at the same time, many projects at the same time. I'm talking about like seven, eight, 10 projects at the same time. [snorts] I'm just like moving like from a task to another as fast as I can comprehend and it's very tiring, but it's still has a lot left. One eternity later. This is what you get when you dedicate an AI to design only. Okay? Where you get to focus on the design, not one shop prompt. Build me an app and by the way, make it look absolutely beautiful. This is a separate task that just says, "Do me a design." Ladies and gentlemen, I present to you in the red corner GBT 5.4 before with a front-end design skill. Okay. This h this looks like my my drinks app very much. I What did you use to design it? Replet. I saw this and I was like, I guess Oh, no. Replet uses 3.1. Yeah, that. So to be honest, it's nice, but it's very open AI looking for me. They have this tendency much experience. They have this tendency to uh do this like extra round corners like the corners are always so rounded. And I I literally had this and I told my I had a very similar design and I told my uh codeex to remove the the serif font and also don't make the corners so so round and don't have like too many corners. But that aside, I think it's pretty decent. So the only thing that really the only thing that really bothers me it's kind of done some weird stuff like doubled up live session and like there's there's a a card within a card here within a card within a you know so many this is like cardelact but the font is quite nice. Um if we go into create workspace uh let me log in. Oh it has like the animations. So, it's got it's got I can create projects and things like this. It's Do you know what it's done though? It's really changed new. Oh, that screwed that up. It's really changed a lot. And what I want to show you here as well is Gemini 3 design. So, this is this is what it creates. Okay. Yeah, this is a bit uh more like edgy and uh unique. The other one did look like my drinks app to be honest a lot. Without interacting with it, I I find it hard to show you. It does feel like very like blinky. Everything's like a like the Matrix kind of like like everything blinks in. This is one of the projects I made uh in the in the app. And then and then this is like the upload screen. very very simple, but it all of the effects felt very glitchy and blippy and and you can imagine all these like sounds going off. And the app the name of the app's green lit as well, by the way. Okay, so I'm going to stick with that. It's green, obviously. And there's, you know, a couple of things going back to back to GPT's uh you know, go. I should have said do not change the layout, you know, cuz it really has added. you saw the difference between this is the project upload screen you know um it really has changed a lot whereas I should have said don't change anything about the product and I was lucky in in the instance of the the Gemini version is that I didn't say that either so this is again I think quite important I'll be making a video about this by the way um about putting these designs up against each other and getting the best out of them and one of them will be build your app but then have it potentially even do like a design review or something like that. So, yeah, I'm I'm pleasantly surprised. I'm not I'm not I don't hate this design. Um a lot more creative than I thought it would be, but you compare this to some of the designs that people are are complaining about. And they're very much um you know, classic AI generated kind of functional design over something that's slightly creative. This feels to me like something slightly more creative. And this is this is raw GBT 5.4. So yeah, this works. This is like I I wouldn't change too much except for the probably the boxes within boxes. And I mean, look at it. Look at this. Isn't Don't Don't you see the similarity? I see similarities just just with Yeah. But you can tell this is AI generated, I would say. Yeah. And this is from the gradients to it. Yeah. This is from uh some time ago and I didn't give it any design direction. This is more about the functionality of the app. Mhm. Uh once you log in fine from that perspective. Yeah, I will I will change it. But I yet to find a design skill or a library that help me. I I want I want something like Shotsien but more creative. Um but I I guess that's a part of it. Like why don't you create your own skill? Why can't you have a know why can't you create your own skill that does your thinking of how you go about design? The things that go through your mind and build a skill and see what that does. I I'm creating a UI library actually. Well, it's for controls, but maybe a UI library for Yeah, but I'm not talking about a library cuz you want every product to be unique from Yeah. Yeah. based on maybe maybe this skill actually takes you through the like keywords, brand keywords or something like that has more of a process behind the unique application of this particular design. Do they want fun? Do they want serious? Do they want bold? Do they want? Do you know what I mean? The classic journey of how you come up with the design. Build that into a skill and have it then run through your design methodology as all as part of that skill. Yeah. I'm I'm going to check also the skill directory and see if there is something like that for it. Yeah. Yeah, definitely. Good place to start and then you can edit it and you know this is part of a larger conversation on my show. Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.