I Tested Sonnet 5.5 vs Opus 5.5 vs GPT 6 Astra (No Hype) transcript

So much for the AI slowdown. These last few weeks have seen some insane models get released, specifically GPT6 Astra, Claude Opus 5.5, and now Sonnet 5.5. And now the question you should be asking is which model should I be using? Because the answer isn't always obvious. You might think that these flagship models like Astra or Fable are going to crush the smaller competition, but that isn't always the case. In fact, when we look at something like Sonnet 5.5, it actually posts better benchmarks at Agenta Coding than its bigger brother, Opus.

But how do these three models compare when we put them head-to-head in realworld tests? Well, that's what we're going to figure out today. And by the end of this video, we should be a little bit closer to answering the question of which model should we use. Now, before we jump into the tests, let's go over some of the benchmarks and the pricing very quickly. And these are just some of the benchmarks where we have data for all three of them.

Now, Opus 5.5 by the numbers wins four out of our five benchmarks. And Sonnet 5.5, oddly enough, actually beats them all in terminal bench 4. And this is the smallest model of the three. When we look at pricing, Sonnet 5.5 is half the cost of Opus at $2 per input and $10 per output. Astra is the flagship of the GPT model. So unsurprisingly, it is the most expensive. It is five times the cost of Sonnet. So when we're taking a look at the results we see today, keep in mind these price points.

Now for our first test, we are going to see how these models do. When it comes to creating these JavaScript style animations, what you see right here was created with cloth. This didn't use any video generation tools or image generation tools. This is pure code and you are able to create videos just like this one. So, I gave all three models the same prompt, telling them to create in JavaScript a 15-second explainer video telling the audience how sub agents in Claude work.

And the first one we're going to look at here is from Sonnet 5.5 and all the prompts and everything you see in today's video. I'll put all those resources down in the pin comment. So, here's Sonnet. So, not bad. I like the audio and the actual visuals kind of matched. Every single beat had something going on in the actual image. Well, in the actual video. Although I will say in some instances it was so fast it was actually hard to see what was happening.

But the fact that this isn't an actual AI video, this is pure JavaScript is really cool. Now here's what Opus 5.5 created. And I told all three of these models to use this sort of handdrawn style. So, I thought that was pretty good as well. You definitely can see the similarities between Opus and Sonnet, but I do think on the first pass I liked Sonnetss a little bit less and would definitely give Opus the edge. And here is Astra.

And GPT6 definitely leaves something to be desired when we compare this to both Obus and Sonnet. Not nearly as dynamic. The Anthropic models when they created these JS explainer videos, they were very quickly paced, almost to a fault, versus this one really slow and pretty boring actually compared to the first two. And the audio also isn't nearly as impressive. One thing that does play in Astra's favor is that both Sonnet and Opus took about 300,000 tokens to complete this versus Astra which took only 70K.

But overall, I would give the edge to Opus 5.5 year closely followed by Sonnet. Now for test number two, we're looking at front-end design. How well can it create a landing page for a boutique hotel called Dune House? Now, the Enthropic models do have a bit of a disadvantage here because they can't create their own images like GPT6 Astra does. So, to help Sonnet 5.5 and Obus 5.5 out, I told them in their prompt they have access to the Higsfield API.

Now, if you haven't tried out the Higsfield API yet, I highly suggest you do because you're no longer stuck paying for their monthly subscription plan where you're doing like 200 bucks a month. Instead, it's simply pay as you go and the API prices are as good or better than anything else out there. So, it's perfect if you're using it with something like Claude Code and you just need to do certain videos and images here and there.

They're running an offer right now with cash back on certain API calls as well as extra discount. So definitely take a look. So right here we have Sonnet 5.5 and I will say that Anthropic lately loves these really big hero images on the landing page. I'm not against it. I think it looks kind of cool. I don't love the font choice to be honest. I also don't love this check availability sort of pill background they have here.

It looks really ugly. But I like this. This looks like something I would totally see on a real hotel website where I can check availability. As I scroll down, the images are kind of front and center, which I like. And we see it down here as well. Again, I hate the font it decided to use. And I do kind of like how as I scroll through, I can easily see these different images. Again, lots of imagery. And I do kind of like what it did at the bottom here near the footer.

This sort of is a call back to the hero section where we have that giant image acting as the background. And overall, I would say this is pretty solid for a prompt that didn't give it a ton of direction. And this is, you know, if we just think of how this did in terms of design a few months ago, we're seeing such a leap forward with the 6 series and GPT and these 5.5 series here with Sonnet. Next, we have Opus. And right away, I like this one a bit better.

I kind of like how it did this dune house having it, you know, sort of transparent. It's sort of cut off. It took some creative liberties, you know, versus Sonnet. This seems like looks cool but pretty standard. Opus kind of did its own thing. Again, I like the track availability here a little bit better than what we saw with Sonnet. This looks ugly as hell. This is much better. Even give it it a cool little logo up here on the top left that actually reacts to my mouse hovering over it.

I really like that. It added a little glass morphism here in terms of being able to check availability and it loves this orange to black thing it's got going on. And oh, as I scroll down, you know, the numbers you saw there, they kind of ticked up. We have a banner which stops scrolling as my mouse goes over it and it added some imagery. It even sort of cropped it which looks kind of cool. And as I click through it again, I can look at specific dates and availabilities for those specific rooms, which is actually really nice.

I like that. As we go down, we got more images. I definitely like doing this sort of cropping style. As I mouse over it, it kind of we have some movement there. Also kind of like a subtle design effect I also enjoy. And as we keep scrolling down at the bottom near the footer, again, similar to Sonnet where it has that big image in the background. So overall, I think you see a lot of similarities compared to Sonnet, but just like better and with more unique design decisions that we don't see with the more simplistic Sonnet 5.5 version.

And here's what we got with GBT6 Astra. I think this big picture in the hero section is definitely just sort of like a theme we're seeing with a lot of these AI design websites. And if this is what AI slop looks like these days, hey, we're not in a bad spot. Honestly, out of all of them, when it comes to the hero section, I like GPT6 Astros the most. And I think it is in general the most clean. I've seen a lot of people kind of like say they don't love the GPT6 web design aesthetic.

And in general, I've really liked what it's come up with. And as I scroll down, I still really like it. And I would say it's better than Sonnet. I enjoy this better than Sonnet, but I don't think it's one of those scenarios where like GBT6 Astra destroys Sonnet and it's on par with Opus. It's very much a subjective thing. I think all of these look really, really good. I think web design when it comes to AI is in a great place because when we take the GPT6 Astra landing page and compare it to the Sonnet one as well as the Opus one, I don't think there's a huge drop off between any of them.

I would say again that I would call Opus 5.5 the winner here, but I don't think Sonnet and GPT6 Astra are far behind. And as for token usage, SA and Opus were around 200,000 and Astra was about 125,000. Now, before we jump into the third benchmark, a quick word from today's sponsor, me. So, inside of Chase AI Plus, I not only have a Claude code masterass, but a codeex masterass as well. And it is the perfect place to go from zero to AI dev, no matter your technical background.

Inside, you also get access to my Aentic OS, which supports not only cloud code, but Codeex as well. So, if you're someone who wants to get more serious about AI, definitely check us out. There will be a link in the pin comment. Now, for our third test, we are again looking at front-end design, but I let it go kind of wild. In this case, I told it to create a 3D travel dashboard web application that looks like something out of a sci-fi film.

I gave it a lot of leeway in terms of what it wanted to create. And I also gave the anthropic models again access to the Higsfield API. So if they felt like they needed images or videos in this product, they could create them. So the first one we're looking at right now is Sonnet. And it went pretty crazy. I actually have no idea what I'm looking at. So essentially, it looks like it's flights. So I have the ability to zoom in.

If I hit something like Paris, that's kind of sick. Looks like it created these images using probably one of the image generators inside of Higsfield. gives me a little rundown on Paris itself. Shows me, you know, hey, here's sort of what it's going to cost you. Here's the temperature and here's where like different non-stops. Um, cool. Let's see if I back out. It has like music going on. I'm not sure if you're able to hear it.

Again, I can zoom in pretty far. Click on Oh, let's see if I click on something like Tokyo over here. Again, pretty cool stuff. Um, if I want to depart from here, I can then click somewhere else like Bali. And now it's showing me some prices to Bali from Tokyo. So, honestly, looks pretty sick. Now, here's a look at what Opus created. Again, pretty similar, although I will say not as like wild visually as Sonnets was, but Sonnets was pretty crazy.

So, I can scroll in here as well, though I can't scroll in as far when I click the different cities. We have sort of a similar thing going on where it pulls them up. So, again, I get got some imagery that it created. And I'll move this over here so we can see it better. I can see what the fairs will be from JFK, how long I'm in the air, all that good stuff. says, "If I hit H, I can redraw the world around my departure city." I don't even know what that means, but okay, that's kind of sick.

Um, very cool. Go back to orbit. Let's do that again. That's kind of cool. It kind of takes the whole globe and just turns it almost like flat. Pretty sweet. And I can see all again all the different cities I could travel to and all those things. switch to LAX or LHR. Now, I kind of like this one better than what I saw with Sonnet. Sonnet looked awesome, but I like sort of the additional things it added here, specifically like flattening the globe versus what I saw in the first setup.

And here's a look at Astra. Now, this one is much more subdued versus the, you know, visual spectacle that was both Sauna and Opus. So, right away, I can choose where I'm departing from. I can put the departure date, the return date, how many travelers, what cabin I'm doing. I have some information down here at the bottom. If I click explore, it brings up the image over here, as well as some more information about where does it actually be going.

And I think right away you see that compared to the Opus and Sauna version, this one definitely feels a lot more practical. You could definitely turn this into something you would actually use versus I think the sauna and opus ones are definitely a spectacle like I talked about before, but you could argue that based on the prompt I gave it, it's definitely preferring form over function. And it's not necessarily Opus or Sonnet's fault.

So in regards to where I would rank all three, I would probably put Astros number one purely because this is a much more functional application. If we're going straight off of visuals, I would put Opus number one, followed by Astra, and I'd put Sonnet last because I think it just went too off the rails. Now, as for token costs, Opus and Sonnet, again, they seem to hover around the same token cost every time. They were about 300,000 and Astro was half that.

Now, for our last test, we're looking at game design. I had all three models create inside of 3JS. So, this just runs on the web browser, their version of World of Tanks. So, let's see what they were able to come up with. So, here's what we got with Sonnet. This is sort of the loading screen. I have a few different tanks that I am able to choose from. And you can see over on the right, there's different stats for every single one of these tanks.

So, let's go with the TD7 tank destroyer. Looks like it creates some bots for us. And here we go. So, I also have different ammo on the bottom that I can switch through with my hotkeys. And again, we created all of this from scratch. I didn't, you know, have it take any asset packs or anything like that. So, let's see what happens when I find some enemy tanks. All right. So, we can see some tanks, I think, on the horizon there.

I can actually zoom in here, which is kind of cool. Looks like I'm getting hit. Has like a built-in reload time. Actually have to repair here. And uh yeah, I got destroyed. [laughter] So uh I'm not very good at this game, but this actually looked like it was working pretty well. Now we're looking at Opus. For reference, Sonnet, which is what you just saw, took 750,000 tokens and Opus took 500,000 tokens. So let's try the heavy tank this time.

You can see like before we have all these stats on the right hand side. We have a little overlap right here on the bottom right. Um, sort of minor, but something to take note of. And let's get started. So, similar setup to what we saw with Sonnet. You're not able to hear the audio, but it's exactly the same. And these two tanks just totally blocked me. Same sort of system as before where I have three different sorts of rounds I can choose from.

And I also have the ability to zoom in and out. We also have a map that is very similar. And let's see if I can actually take out another tank this time and not die. Uh, I think I'm going to die again. Yep. Second time was not a charm, but it worked pretty much identically to Sonnet. And now, let's see what Astra gets us. Just like we had with Sonnet and Opus, we have four different options. Light, medium, heavy, and the tank destroyer.

We don't see as many stats on the right hand side as we saw with the anthropic models. And overall, it's a bit more of a simplistic user interface, but let's see how the actual controls work. And again, for reference on the token side, this took about 300,000 tokens. So, camera is a little bit more sort of sensitive, but I'm not against that. I kind of like the map here a little bit more, and in general, I actually like the UI it came with a little bit better.

Although, I probably will do just as poorly as I did in the last one. Let's see if I can actually get a kill. Oh, we might actually get somebody. Oh, one of my teammates got him, but I'll I'll take it. And what's also cool about this is it has like, you know, it actually is able to like tell the angle at which you get hit and all that. So, it'll tell you if it's a ricochet and it depends if you get it from the front or the side and stuff like that, which is kind of neat.

But overall, I would say with these three, pretty much the same. Not a huge difference between them besides some minor user interface stuff on the Astra side, but Opus and Sonnet were basically the same exact product. So honestly, I would call it a tie. You could maybe argue was an advantage to Astra because it used way less tokens. Sonnet definitely took the most tokens and took the longest to complete this. So overall, when we take all these benchmarks into account, I think the winner here is Opus 5.5.

But the other two models, Astra and Sonnet, are not far behind. And that's huge news if you're an anthropic user because sonnet is so cheap and sonnet 5.5 is such a massive upgrade compared to sonnet 5. Sonet 5 was basically unusable. Same with opus 5. And if anyone thinks this is me like disparaging open AI, I'm definitely not because I love Astra. I use it literally every day. And this competition is awesome for us because I don't think we'd be getting this quality of outputs if OpenAI wasn't forcing Enthropic to be better.

There's a reason everyone was loving Astra a couple weeks ago because it's a great model. But at the end of the day, you need to try these things out for yourself and see what works for you. But again, you have options. So that's where I'm going to leave you for today. Hopefully you got something out of this. As always, make sure to check out Chase A Plus if you want to get your hands on my Claude Code and Codeex Masterass.

And besides that, I'll see you