Astra vs Fable 5.1: Which AI Builds Better Websites?

Igor Pogany tested Anthropic's Claude 3.5 Sonnet and OpenAI's GPT-4o on building websites from single prompts across 50 diverse categories. Astra won on visual quality and professional layouts, while Fable 5.1 excelled at creative interfaces. Astra was also cheapest at $0.41 per site.
Chapters
AI is really, really good at building websites. Whether it's tools for yourself or others, learning experiences, something that you're using instead of a PowerPoint presentation, or just a prettier way to look at AI output. Websites are one of the big capabilities. And I think you already know that. And if you don't, by the end of this video, you'll surely know this. But with all these new AI models coming out, which one builds the best websites? Is it Claude's Fable 5, 5.1? Is it the new GPT-6 Astra? We by now know that Astra can do some crazy stuff. And I covered that on this channel. But concretely, I just wanted to know when it comes to creating sites, which is the best one. So I created my own benchmark where I created 50 sites from different categories, and I generated them with these different models. And then I blindly rated them myself. So it took some time, it cost quite a bit of money. And in this video, I'll present to you the result of all of this AI website frontend building, whatever you want to call it, to ultimately answer the question of which AI should you be using to build your websites and which model is good at what. In the end, we'll have a winner. But let me tell you that the nuances matter because in some categories, the one is better, in other categories, the other. So let's break it down. So my curiosity started with asking, okay, with these new Anthropic models, which one is best? Historically, Anthropic used to be the frontend building king. Everybody agreed that the websites from Claude were just the prettiest and the best, especially when you one-shotted them. And that's what all of this is, by the way. So let's talk about the methodology here briefly, because all of these are one-shot prompts, meaning prompts for various websites that you just put in, and we rate the first output. We don't go in and refine and make things better. It's just what you get once you request something. And I started out with a test set of 50 sites that I created in Fable 5 and 5.1. But then I quickly realized there's not enough variety amongst them. Yes, they were different websites, but a lot of them were like personal websites or ecommerce storefronts. And I wanted some of the funkier stuff in there too, like this Murmuration Field Lab where it animates bird flock movements, or this Ink Current Studio where you can paint and it makes these visuals, creative web stores like this coffee configurator, and indie game lab launch sites. There's a lot of variety in the second test set. And that's what we'll be focusing on. And that second test set, these 50 sites, 50 different prompts that I gave to the model straight to the API, no other context involved. Well, they exist across 10 categories. And each one will be rated for 3 values. First, and most importantly, my own blind judgment. Just how much do I like it visually, functionally? What is my preference between the 2 sites? I really took my time in rating these. And I did that blindly for all of these in different matchup combinations. Combinations, like Fable 5.1 versus Astra, etc. Secondly, what does the AI think of the visuals? AI blind judge looking at the visual quality of these sites. And thirdly, AI blind judge on the functionality. In other words, AI looked at, hey, everything that we prompted for, is it present in the result? And is it all fully functional? Across these 3 rubrics, we'll be making a judgment on what is the best website generator from all these models. And hey, if you're out there ready to type your comment, Igor, this is not a scientifically valid experiment. Your qualitative judgment is not rooted in the scientific methodology. Yes, you're right. I was just curious which one I want to use for websites moving forward. And I created a benchmark that satisfies my requirements. Hey, if that's not good enough for you, feel free to come up with your own and share it with me. I just want to know what fits my taste and if it's functional. And I figured that'd be useful to you too. So let's begin, starting with Fable 5 versus 5.1. And I'll keep this segment brief because I think a lot of people really want to hear the Astra stuff. But last week, as you might know, Anthropic released Fable 5.1 to get ahead of the GPT-6 Astra launch. And it's an amazing model. Benchmarks were really good. And the sites that it creates are different from Fable 5. Let me just tell you the result of this. Fable 5.1 across the board was just better at these websites. Now let me tell you some nuances of this. But I think the biggest one is the spacing on these sites is just better. I'm not going to show you all 50. I will share this link with you if you want, you can go in, all these 50 sites are hosted on the web, you could click around, try things yourself if anything catches your eye, or, you know, feel free to use them for inspiration. But this is the first thing that I noticed. Look, the way things are spaced out with Fable 5.1 here on the right, it's just better. It has a bit more breathing space, it makes more sense. And there's like a visual hierarchy present. In the site. Whereas here, I don't know, this text is the same size as this, as the color pickers, as are the buttons. Whereas here, this headline dominates, this is cut off at a reasonable point at the bottom. It just makes more sense design-wise. And I found this to be true across the board. Additionally, I found that the SVGs were way higher quality with Fable 5.1. And it just expressed a bit more creative taste, I felt like. I mean, look at this line. These are the same prompts, All across the board, right? And again, here, the spacing, look, it's not overwhelming, this interface, where this one on the left is like, it's too much. So let's get to our 3 rubrics and move on to Astra. But basically, Fable 5 versus 5.1, on my human preference blind test, Fable 5.1 won 30 out of 50 blind tests. And 17 of those actually went to Fable 5. You know, so it's better, but it's not like every time I would prefer Fable 5.1. And there were 3 ties, if you're wondering what happened to the last 3. Okay, so I clearly have a preference for Fable 5.1. I gave you some of the reasoning. What did AI think of the visuals? Well, it said that Fable 5.1 won 40 out of 50 times, Fable 5 only won 7 times, and it gave it 3 ties. So it also agrees that Fable 5.1 is better, but it was even more extreme in its opinion. And when it comes to functionality, they were actually dead even. Everything functioned exactly as expected. So in terms of functionality, it didn't really make a difference. So clearly, if I have a choice between these 2, I'd be choosing Fable 5.1. Alright, so that makes that the winner. And now let's look at what many people have been waiting for, like Fable 5.1 versus GPT-6 Astra. Which one of these is better at making sites? And here we'll go into more detail because the nuances matter here. So first, some of my unstructured thoughts for you. So clearly, the SVGs, the visuals that this generated with geometry, were always superior with Astra, as you can see in this example. Or you can really see it if it tried to recreate the world map with vectors. Here, clearly, Astra is just better at drawing. A second big theme that I noticed is that Astra is just more professional. And Fable 5.1 is more cute, for the lack of a better word. I mean, it really is. Look at some of these examples. Every Astra generation looks a bit like a SaaS. I mean, look at this trip packing planner. This looks like a SaaS product with the header, the navigation at the top almost every single time. And then the trip packing planner, well, it looks like something a teenager made, which, you know, is not bad. It's a bit more casual. I kind of like it. It's a trip packing planner. Or about the site that promotes a sketching cities course. Well, here's the Astro generation. There's some funkiness here with the window. That's undeniable. But again, you have this professional site, you have this highlighter that has been used well. Everything is spaced really well. And this one that was made by Fable, it's just a bit more cute. Or maybe you could say that it has more personality. I don't know. But it's definitely different out of the box. Now, sure, you could steer all of this with your specific styles and guidelines or examples that you give it. But hey, this is the default behavior of the models. And oftentimes you just make a site for a single purpose. And this is what you'll be getting. Another thing that I noticed is that every game that I created from a one-shot prompt was not even close. Astra crushes games pretty much every single time. We'll review the detailed stats here in a second. But man, look at this Depthcrawler made by Fable, which, you know, it's an icon. And you somehow have this map and you move around. And fair enough, this has always been kind of one of those classic games that you make with AI. Look, I think that's a coin. This is an enemy, that bee. All right. Now look at what Astra made. Same prompt, by the way. The enemies look like actual creatures. It's chasing me. I got it. I look like an actual character. It's a whole thing. Visually, it's not even close. This goes for all the games. And I'm brushing over some of these because, hey, if you're curious, you can just try these yourself. The link to this site with all the sites will be below, no strings attached. If you're enjoying hands-on tests and real examples like this video has, well, that's what I do on this channel. Every week I upload new videos, so subscribe to get those. And now let's have a look at what we have here next. So those are some first thoughts. So what did the final rating look like between Fable and Astra? Well, when it comes to my personal rating, the human preference blind choice where I opened every site and I looked at the different views, mobile, mobile, web, I tried the application. Well, GPT-6 Astra won 35 times out of 50. And Fable 5.1 took it 15 times. So a strong preference, 70% of the choices went to Astra, if you ask me. Now, here's an interesting thing. If you ask AI, it was so opinionated on this. 47 out of 50 went to Astra when you ask AI. And 3 of them were ties. Not a single time did it say that Fable did a better job when it comes to the visuals and the look and feel of the whole thing. And I was really surprised by this. I was like, wow, not a single time? Like, really? I had 15 cases where I preferred Astra. You had 0 AI? So I'm like, hmm, this must be because with my agent, I'm using GPT OAuth, and I'm using the OpenAI models to rate this. So like, ah, okay, it really likes its own style. That makes sense. So I reran the whole test here. And let me tell you, I'm just reading off these examples here. But when I used Fable 5.1 as the blind judge on these, 48 out of 50 went to Astra. And Fable 5.1 won 2 of these. And then there was 0 ties. So it had the same results. But yeah, AI really, really prefers the Astra generations in terms of visual appeal to human beings, for whatever that's worth. And then finally, the functionality, because this matters if one of them works better than the other. And there's certain buttons or functionalities within the sites, like there's a lot of interactive apps here, like music synthesizers and situation and physics simulators, educational apps. In terms of functionality, GPT Astra passed 48 out of 50 tests and Fable 5.1, 47 out of 50. So pretty much the same, which as a reminder, Fable 5 also got 47. Curious which ones Astra actually failed on. And it was these 2, the AI request topic bubble explorer, and this Terminus app, which basically simulates a personal terminal in the web. And yeah, this thing has a lot of functionality. And it's quite complex. And Fable one, one-shotted this successfully, whereas Astra had issues. Whereas Fable failed on this Ink Current Studio. See, this doesn't even draw if I go through. Whereas the Astra version, yeah, it just works. Change these values to different color, export the still, all functional. And then also this RAG learning app and this ocean restoration annual report site also had malfunctions with it, although it's really simple. And hey, by the way, if you have any questions about anything that I talk about here, or you want more definitions or details on something, just leave a comment below. I'll go into some of the top comments and answer them myself and give you all the details you might want. Try and keep this video pretty high level so everybody gets value from it. Also overall, as you might have expected, you know, it's a pretty significant win for Astra, both of my preference and the AI preference. And functionality is essentially the same across them if you're one-shotting frontends. But this leaves us with 2 questions that are unanswered. One of them is, okay, so which categories of apps performed better in one than the other? And I have some answers for you. And 2, how much did all of this cost? Well, let's have a look. So with the category preferences, I'll focus on Fable 5.1. As that was the winner on the Anthropic side versus Astra. In terms of my human rating, I preferred Astra on all of them except of audio and music interfaces and experimental interfaces. Some of the creative and out-there stuff that requires more of a different type of personality did well with Fable. Just like I said in the beginning, I mean, that's the strength of the Fable model. It's just a bit more, I don't know, opinionated, a bit more unique. If you will. So for example, look at the sites where it creates a kinetic poem. It just looks very corporate on the Astra site. Whereas with Fable, well, look at this animation and this entire site that was made by Fable. It's just, you know, it understood the assignment. It's not bad either. I just think the other one fits more. On the other hand, where it was a clean sweep for Astra, where I just clearly preferred it every single time, 5 out of 5 times, I guess, was this category, narrative, editorial, and spatial. Just these funky sites that kind of tell a story or are supposed to express something in a creative visual way. Like this roasted tomato recipe with a built-in tomato timer and portion sizing. And then when it comes to this museum exhibition guide, you can just see it really did a great job at creating these images. I mean, this chair is not perfect, but otherwise, it's very visually appealing just with SVGs. Of course, this is what Fable created, which is just, I don't know, simple. I prefer Astra. And then I think you'll like this one as a recommendation. When it comes to dashboards, it's not even close. Astra every single time. Data visualizations, dashboards, just looks better, feels better. Look, this is the Astra generation of dashboards for the economic index right here. You can, you know, expand them, contract them. And it's just a good layout. Whereas this, I don't know, it feels a bit misspaced. Or this global AI usage map generated by Astra. Just a beautiful dashboard with everything working as expected. And again, the SVG generation is better, but let's not judge it just on that. Whereas here with Fable, you get this little scrollbar that to me doesn't really fit into the whole thing. And then this is very text heavy up top, where this is a nice structure to it. I mean, one of these more professional things. Yeah, Astra seems to be the way to go there. And then lastly, let's talk about cost. What did all of this cost? Well, let me tell you, in putting together this benchmark, I spent well over $200 just testing different prompts and seeing what I like. And again, it took several iterations and I threw out versions of this before this one. But when generating these 50 sites directly through the API, no caching, no batch API, just straight up give me these sites from the model provider, no extra context with it. Fable 5 cost $21.24. Fable 5.1 was more expensive at $29.15. So that's almost 60 cents per site. And GPT-6 Astra came in at $20.64 for all sites. Costing around $0.41 per site. Meaning that it didn't just pass the visual vibe check of mine in 70% of the cases versus all the other models, but also AI agrees with that opinion. And in terms of functionality, you're not going to see much of a difference when one-shotting sites. And in terms of cost, it's the cheapest too. So there you go, browse some of the sites. Let me know if any example in here inspired some idea for you, or if you have further questions. If you want some examples of what Astra can do in practice, check out my channel. I have some amazing videos on that. People seem to love those right now. And with With that being said, if we have not met yet, my name is Igor. I look at this stuff and figure some things out for you so you don't have to do the work. And with that being said, that's pretty much everything I have for today. My name is Igor, and I hope you have a wonderful day.










