Claude Just Killed Opus (NEW Sonnet 5.5)

0:00Anthropic just released their next generation Sonnet 5.5 [music] and the benchmarks are absolutely insane. In most tests, it's right up there with Opus and in a couple, it actually beats it. So, to test their claims, I gave both models the exact same prompts on four real jobs starting from [music] animated ads to playable 3D game. And by the end, you'll know exactly if the benchmarks hold up or not and which Claude plan is worth paying for. The benchmarks on screen right now come from Artificial Analysis, an independent site that runs every new model through the same set of tests. On their overall score, Sonnet 5.5 gets a

0:3956 and Opus 5.5 gets a 58. So, the difference between them is only two points. On agentic coding, which is the model writing and running code on its own, Sonnet actually wins [music] with 63.6% compared to Opus's 59.6% and on automation, it's a tie with Sonnet at 71% and Opus at 70. So, the first job I'm giving both of them is a coding job you can actually watch. A 30-second animated ad. [music] They run in Claude code, which is the version of Claude that can work with files on your computer and they build that with Remotion, a free tool that turns code into a real video file. I'm not helping

1:18either one, so they have to install everything, write code and render the video themselves. And if you want the prompts I'm using in this video for yourself, there will be a free resource below where you can grab them. The first ad is for Piranesi, a novel by Susanna Clarke and the only reference each model gets is a photo of the cover. After a while, both are done and here are the two videos. Starting with Sonnet 5.5, it got the fonts, the colors and the themes on point. But as someone who has read the book, the ad doesn't really make sense and I don't like the sequencing

1:50Sonnet went for here. And even though the whole thing is smooth and looks good, I would like it to be much more detailed. Then there's Opus. Right [music] away, the figures in its 3D hole have a lot more detail, and I especially like the cinematic camera movements it uses [music] to sequence that. It gives us a smoother transition while selling the book. So, I definitely like this one better than Sonnets, but honestly, that one wasn't too bad either. [music] So, on this first ad, Opus takes it, but Sonnets isn't far behind. And if you want to go from casual [music] using ChatGPT to using AI like the top 10%,

2:23that's exactly what my AI fluency community on School is for. There's a new video every day for 12 weeks, and the price goes up as more people join. So, the link's down in the description if you want in at the current rate. Next, I wanted to see how both of them would handle something real, a motion graphic ad for this exact School community. [music] So, to do that, I'll follow the same steps again, and this time, the only image they get is a screenshot of the

2:45School page.

3:37Sonnets goes first this time, and it got the colors right with pretty smooth animation. It's nothing extraordinary, but it's on par with what you'd expect. And with Opus, the gap is a lot narrower this time because the two look almost the same, and there's not much to tell them apart. So, Sonnet definitely got super close to Opus here and that's impressive for a model like that. These are just test runs though. If you're wondering whether AI can make motion graphics for real work, it can, but you have to be a lot more careful with your process. [music] Tell it exactly what you want from the fonts, the colors, and

4:10the motion, each one separately. Anthropic's own guide says Claude falls back on safe generic designs unless you do. Give it real reference to work from a brand guide or a few examples you like along with your own input. And if you're doing this in Claude code, install emotions official skills first because they teach Claude how to handle timing,

4:28fonts, [music] and rendering properly. So, with that much detail in the prompt, either model can give you something you'd actually post. And even though motion graphics are impressive, [music] they're not something most of us make every day. But the one thing a lot of us do every day is write with Claude. And [music] to test that, I'll make both models go through a structured process to write a dummy script. The process itself is pretty simple. First, it has to research my audience from one vague line. People who use ChatGPT or Claude a few times a week at work but feel like

4:59they're only using about 10% of it. Then, it has to find what's trending for that [music] audience in the last 30 days and pick one topic by itself. After that, it writes a 1600 word script on its own following a list of my writing rules like no m-dashes, no hype words, and no made-up facts. Both run in the Claude desktop app in co-work [music] mode and I'm not answering a single question along the way. So, everything has to come from the one big prompt I gave them. [music] And this is also where the benchmarks say they split. On real-world office work, they're basically [music] tied with Sonnet at

5:321844 and Opus at 1846. But on factual knowledge, Opus scores 66% against [music] Sonnet's 54. And when Sonnet doesn't know something, it's less likely to make up an answer 47% of the time against Opus's 59 to be exact. So, here are both scripts side by side. So, next topic is workers taking their AI skills. Opus went with difference between the chat and work tabs in ChatGPT. And honestly, Sonnet wins the topic by a big margin because it's a much more likely fit for my audience than what Opus gave us. The hook is a lot better on Sonnet's side, too. While Opus's [music] hook and topic are both

6:07pretty confusing. The writing itself isn't that different, though. And either script is pretty good. The one detail that stood out to me is the end of its script. I never told either model what my next video is. Sonnet left a note asking me to fill it in. While Opus made one up. A video comparing ChatGPT and Claude that I never planned. It's only one example, but it matches the benchmark where Opus knows more and Sonnet is less likely to fill a gap with something that isn't true. Whichever model you use, double-check anything it says about you or your plans because it has no way of knowing those [music] and

6:39it might just guess. So, surprisingly, on picking the topic and writing the script, Sonnet beats Opus. But the script is still a job with a clear finish line. And Anthropic itself says Opus is still clearly stronger at complex, open-ended work. So, for the last test, I'm giving both of them the most complex job I could think of. A playable 3D game set in Rivendell from The Lord of the Rings. They have to build a valley the way Tolkien describes it in the books. Then, they turn it into a game with original characters and a quest [music] to light seven lanterns

7:12before sunset. All inside the browser. All of it is free. [music] Three.js for the 3D, textures from Poly Haven, and character models they were styled themselves. [music] I'm backing Claude code for this one. And this time, each model runs a whole [music] team of helper agents at once. Sonnet uses nine and Opus uses seven with one on the water, one on the characters, one on the quest, and so on. Both build the core of the game themselves and never let another or touch it. But they run their teams differently. And keep in mind, running this many agents at once burns through your usage fast. [music] And the

7:48end result is a single file you just double-click to open with nothing to install. Let's start with Sonnet's version. It opens on this cinematic drone shot of Rivendell, [music] and you can see the controls and the graphic settings at a glance. Personally, I'd go with high, so I'll select that and click begin. I start on the mountainside, so I'll hold down shift and W and spring down through the valley. The forest loads as I go, and honestly, it's pretty good. I like how Sonnet handled the textures and the mix

8:16of all these different kinds of trees. [music] You'll run into a lot of characters along the way, and everyone is a free model restyled for this world. And [music] if I walk up to this gate, I can talk to the gatekeeper about my quest. The water [music] is really impressive, too, and I love how clear it is. The house it built is pretty nice with the medieval feel we know from The Lord of the Rings. Behind it, there's a stream coming down the mountain, and even though it's not super realistic, I like the style Sonnet went for. I'm fairly impressed by how it handled a map

8:47this big because I can walk a long way out, and there's still more to see. Next up is Opus, and both models took roughly the same time to build their games. The opening is close to Sonnet's, so I'll click begin the journey. Right away, everything looks different, and [music] I spawn somewhere else entirely. From the textures alone, I can tell Opus has gone for a much more realistic [music] style than Sonnet. So, I'll spring down to the lower valley of Rivendell and explore the rest. The terrain here is rougher, where Sonnet's was more cartoonish. [music] And Opus planted about 3,500 trees across five kinds, with pines [music] on the high slopes

9:23and oak and batch down in the valley. At the river, the water is clearer [music] and more lifelike, too. I really like how the stream steps down over several levels here, though it's a bit glitchy. The house and the other buildings are a lot sleeker and more structured, and for some reason, I can't go inside. And that's the biggest difference between

9:43the two. [music] Sonnet has gone for a more traditional look, straight out of Tolkien's books, while Opus has gone for a more modern touch. Opus does follow the books' layout more closely, though, with the zigzag [music] path down into the valley, the narrow stone bridge with no railing, and lots of small waterfalls. I still prefer the traditional feel of Sonnet, [music] but Opus wins the one on realism because of its textures, its water, [music] and how it shaped the terrain and the architecture. So, the hardest of four jobs goes to Opus, but Sonnet still built a complete game from one prompt, too. But the working game

10:19doesn't tell you yet which of these two is worth paying for. And that [music] question comes down to what you actually use Claude for. So, do the benchmarks hold up? Mostly, yes. Across the four jobs, Opus did better on the Piranha game, Book Cat, and the Rivendell game. The School At was basically a tie, and Sonnet won the writing test. And every time Opus won, Sonnet wasn't far behind, which is what two-point gap on the benchmark should look like in a real work. The one claim that didn't show up for me is Sonnet's lead on agentic coding, because it never beat Opus on a complex build. But the part about Sonnet

10:57making things up less did show up [music] when it left my next video blank instead of inventing one. So, here's how I would pick. If you use Claude mainly for writing, research, and everyday questions, Sonnet 5.5 is more than enough. And on top of that, you can use [music] it on a free Claude account. But if you build things in Claude code, you're already paying, because Claude

11:20code only comes with the paid plans. [music] And as you saw with the ads, Sonnet gets you through most of those builds just fine. For the big, open-ended projects where the details matter, like the camera we built from one prompt, Opus is still worth switching to. It's included in the $20 [music] pro plan, and one thing to keep in mind is that the benchmarks were run at the highest effort setting, while the Claude app runs on medium by default. At that level, artificial analysis found that Sonnet costs more per task than Opus, $7.60 against $5.98 to be exact. [music] So, if you're paying through the API,

12:00running Sonnet on a lower effort setting is how it saves you money. So, in the end, the benchmarks mostly hold up, and for everyday work, Sonnet is the one to use, with Opus saved for the complex builds that need real judgment and nuance. And picking the right model only helps if you're using it the right way. Most people still use these tools the way they did back in 2023, and I made a whole video on what to do instead. So, click the video on screen right now, and

12:28stop using ChatGPT like it's 2023.