I Tested Claude Opus 5.5 — The Results Are Insane
Read the Portuguese translation →
Author: Code A Program · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: I Tested Claude Opus 5.5 — The Results Are Insane · Channel: Code A Program · 16:02 · Watch on YouTube
Topics: AI agents, Software engineering with AI, AI content generation, AI models
Summary
Section titled “Summary”The video tests Claude Opus 5.5 on a series of increasingly complex visual and coding tasks, from SVG generation to interactive 3D scenes and websites. The presenter praises the model’s results, especially its detailed house recreation from a reference image and its cinematic headphone product site, while noting some imperfections in the house walkthrough controls. The final test uses Mixamo characters and animations to build a playable 3D fighting game. The presenter also discusses Opus 5.5’s pricing and compares effort modes, noting that extra-high effort can improve quality but costs more.
Key points
Section titled “Key points”- Opus 5.5 is presented as significantly cheaper than Fable 5.1, with prices of $4 per million input tokens, $20 per million output tokens, and 20 cents per million cache reads.
- In an SVG test, extra-high effort produced a detailed raccoon eating watermelon; the presenter found it much better than the high-effort result.
- A 3D weekly pill organizer included individually hinged and animated caps, varied pill shapes, and keyboard controls.
- A 3D scissor lift featured adjustable height, automatic camera zoom, movement indicators, and a blinking light.
- The house recreation included room selection, lighting options, and a walk-through mode; the presenter noted that mouse control was difficult and some details differed from the reference.
- The Aura Audio Co. website included rotatable products, color options, an interactive headphone disassembly view, and a frequency response chart.
- The presenter reports that Cloud Chrome tested the website across three mobile screens simultaneously.
- The final project was a 3D fighting game using Mixamo characters and animations, with character-specific actions, multiple modes, difficulty settings, and a post-fight results screen.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Introduction and pricing | The presenter introduces the Opus 5.5 tests and reviews the model’s token pricing. |
| 01:02 | SVG and simple 3D tests | The video shows an SVG raccoon, a seven-compartment pill organizer, and an adjustable 3D scissor lift. |
| 05:11 | Recreating and exploring a 3D house | Opus 5.5 recreates a house from a reference image and provides room selection, lighting modes, and a walk-through. |
| 08:16 | Interactive headphone product website | The presenter explores a 3D audio product site with rotatable headphones, interactive components, and product specifications. |
| 11:20 | Building a 3D fighting game | The presenter explains how Mixamo assets were used to create a fighting game with selectable characters and combat controls. |
| 13:22 | Gameplay and final results | The video shows a fight against the CPU and concludes with the game’s match statistics and the presenter’s sign-off. |
Quotes
Section titled “Quotes”Opus 5.5 costs $4 per million input tokens, $20 per million output tokens, and just 20 cents per million cash reads. — 00:30
This is probably the best raccoon I have seen from an SVG generation test. — 01:02
This is the best 3D house recreation from an attached image that I have seen so far. — 05:11
The hero section alone is a 10 out of 10 for me. — 08:47
And after completing the fight, we get a proper result screen showing the rounds won, hits landed, best combo, damage dealt, and attacks blocked. — 15:31
Concepts and entities
Section titled “Concepts and entities”Claude Opus 5.5, Fable 5.1, DeepSeek, Terminal-Bench 4.0, Three.js, Mixamo, Claude Chrome tool, Aura Audio Co.
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Hello everyone. A few days ago, Claude Opus 5.5 was released and this model is seriously impressive. It has shown some very strong results compared to models like Fable 5.1 while also being significantly cheaper. Based on the pricing, it is roughly three to four times cheaper than Fable 5.1. In this video, I’m going to show you some of the tests I ran with Opus 5.5 starting with some quick and simple SVG tests and gradually moving into more
00:30 complex tasks. And at the end, I will show you the real 3D fighting game I built with it using actual 3D characters and animations. So, make sure you do not miss that one. But before we start, let us quickly look at the pricing. Opus 5.5 costs $4 per million input tokens, $20 per million output tokens, and just 20 cents per million cash reads. The input and output prices are also lower than Opus 5. So, it is really nice to see the price coming down while the model is getting more capable.
01:02 All right. Now, let us start with the test outputs. For this first test, I wanted to start with something simple and then move step by step into harder tests. So, the first test is an SVG generation test. I asked Claude Opus 5.5 to make me an SVG of a raccoon eating a slice of watermelon and I used the extra high effort mode. And this is the output I got. This is probably the best raccoon I have seen from an SVG generation test. It is really, really good. I can easily recognize the raccoon and you can
01:33 clearly see the watermelon and the body. Even the hand placement and the eating pose came out perfectly. Some models can fail on these details, but Opus 5.5 handled them really well. You can also see the background, the picnic mat, the grass, and even some of the watermelon seeds below the raccoon. I also posted this test on X to show the difference between the high and extra high effort modes. So, on the left side, we have extra high. And on the right side, we have high. And the output difference is huge. With extra high, I got a much more
02:04 detailed and realistic-looking result compared to high. According to the Terminal Bench 4.0 results, high scored 64.2% at around $3.88 per attempt. With extra high, we get an extra two percentage points in the score. But the cost is roughly two times higher. But that does not mean you always need to use high or extra high. Medium effort can also be a very good option when you want a better balance between quality and cost. And now, let us move step by step into
02:34 some harder tests. For the next test, I wanted to make things a little harder. I asked Opus 5.5 to create a 3js 3D model of a weekly pill organizer with seven compartments from Monday to Sunday, each with its own hinged cap, and the ability to open each cap individually. Again, I used the extra high effort mode. Let us quickly see the result. This is the pillbox I got, and this is also a really good result. When I refresh the page, we can see some smooth animation on all the caps.
03:06 The weekday labels are also nicely done. We can even rotate the entire pillbox 360°. And when we move around to the backside, you can see the shadow appearing properly. Also, look at this thin rod attached to the cap. That small detail makes the hinged mechanism look much more realistic. We also have some cap animation when we hover over them. Now, let us open the caps one by one. And here, you may notice something interesting. We get almost all the different pill shapes and categories.
03:37 This is actually the first time I have seen this many different pill types generated in one of these tests. When I tested DeepSeek, it provided only two types. But with Opus 5.5, we get a much wider variety of pill shapes. We do not get any sound effects, but we do have a close all button. You can also control which caps are opened from this menu at the bottom. And you can even use the keyboard by pressing the number keys from 1 to 7. Overall, this is a really good result. Let us move on to the next one. I asked Opus 5.5 to create a 3.js 3D scissor
04:11 lift with a slider to control the lift height. And this is the output we got. One thing I noticed immediately is that Opus 5.5 is willing to add a loading animation first. We also have the height shown with the scale, and we can control the lift using the slider. Moving the slider increases or decreases the lift height. Another nice detail is that the camera automatically zooms in and out based on the height of the lift. Only a few models provide this kind of option, and I think it makes the result easier for users to see.
04:41 At the bottom, we also have three things showing the current height, the arm angle, and the ram stroke. I really feel like the lift movement is much smoother, especially when it comes back down. And there is one more thing I noticed for the first time in these tests. There is a light on the lift, and it turns on when the lift is going up or down. It also blinks during the movement. I think that is a nice new detail. Overall, I really like this result. And if you look at the material of the platform where the lift is standing, that also looks quite realistic.
05:11 Now, let us move on to some more difficult tasks. For the next test, I wanted to make things much harder. I asked Opus 5.5 to create a realistic 3D model of a modern house using 3.js. I also attached an image of the house that I wanted it to recreate. The prompt also asked it to let us walk inside the house with a realistic walk-through experience. So, this is the output we got. Honestly, I need to repeat this one more time. This is the best 3D house recreation from an attached image that I have seen so far.
05:41 I am not saying it is 100% the same as the reference image because recreating every single detail accurately is a really hard task, but Opus 5.5 almost gave me the best result I have ever seen. If you want, you can also check my Deep Seek or other videos and compare the results yourself. For example, look at this sofa. In the original image, we do not have pillows on the sofa, but in the generated output, it added some pillows. So, there are a few mismatches like that, but overall, this is still a great result. And the best part is that we can actually walk around inside this house.
06:12 I will show that at the end, but first, let us check the other options. Here we have some display options. We can turn the cutaway walls off, so we can see the full walls around the house. We also have an evening light mode. Look at this. It basically switches the house into a night mode. And all the lights and lamps inside the house turn on. That is a really nice detail. And look at this monitor in the room. It is not just a blank black screen. It actually shows a coding screen effect. I really like these kinds of small asset
06:43 details compared to my Deep Seek test. At the bottom, we have an option to press the V key to enter the walk mode, and we can press the number keys from 1 to 7 to jump directly into particular rooms. We can also control everything from this sidebar. So, let me click through the rooms one by one. This is the hallway. Next, we have the living room with some nice sofas and plants. Then we have the kitchen. Next is the bathroom. This room is not as accurate compared to some of the other rooms.
07:13 Then we have the kids’ room with books and other details. Next is the master bedroom. And finally, my favorite room, the home office. Now, let us actually walk inside the house and see how the home tour works. I’m going to press the V key to enter the house. Let us go. I’m using the W key to walk straight, and the movement is really clean. I’m not disturbing the assets inside the house while walking around. I can also look 360° using the keyboard or by moving the
07:45 mouse. What do you think about this result? Let me know in the comments. Now, let us go to the kitchen. We have a coffee machine and other items that we would normally see in a kitchen. And on the opposite side, we have another room. You can see that the door automatically opens when I approach it. This is another room, and I can even run using the shift key. I’m going to speed up this part of the video. Now, one thing I really do not like is the mouse control. It is harder for me to handle, so I am mainly using the keyboard controls.
08:16 That part is a little disappointing. But, there is one thing that really impressed me. We can actually go outside of the house. This is the first time I’ve seen this option in one of my 3D house recreation tests. So, let us walk around the outside of the house and explore it from there as well. Overall, this is a really impressive result, especially considering that the entire house was recreated from a single reference image. For the next test, the task was to build a premium product website for Aura Audio Co.
08:47 with everything created programmatically in 3js. The prompt was basically to create a cinematic headphone product site with a lot of interactive 3D elements. And this is the result I got. Honestly, Deep Seek was the best result I had seen before making this video, but this one goes beyond that. The hero section alone is a 10 out of 10 for me. And what is even more impressive is that the headphone model itself looks like a proper 3D product render, even though the entire thing was generated through
09:18 code. Below that, we have the our story section. This is not the most important part of the test, but it is still nicely designed and fits the overall premium style. Next, we have the product collection, where we get two different headphone models. And here, we can actually rotate the products by dragging them around. This is another really nice interactive section. We also have different color options for the products. Look at how smoothly the headphone changes color when I switch between them. That transition is really nice. I personally like the first product more
09:49 than the second one. So, let us add this one to the back. And we can see the bag count update when the product is added. But now we are getting into one of the most interesting sections. Here we can explore the individual parts of the headphone. For example, when I select the headband, the leather part gets highlighted. We can do the same thing with the other components. And each part has a short description explaining what it is. And this is the really interesting part. We have a slider that lets us disassemble and reassemble the headphone. As I move the slider, the different
10:20 components separate from each other. And when I move it back, everything comes together again. When I fully expand it, we can see all the internal components clearly. And even while it is exploded, we can hover over the individual parts and see them highlighted. Let us assemble it again. Next, we have the measurement section. A lot of other models would probably just give us a simple comparison table. But here we actually get an interactive frequency response chart along with detailed specifications for both
10:50 products. And finally, we have a clean and minimal footer. The entire website is also responsive and all the sections work nicely in the mobile view. There is one more thing I noticed and I actually shared this in my X post. When I tested the website using Cloud Chrome, the browser interaction looked seriously impressive. Instead of testing the responsive layout on just one mobile viewport, it opened three separate mobile screens and tested the website across them at the same time.
11:20 That is something I had not seen in my previous tests. And it looks like the browser interaction itself has improved a lot. So, this was another huge test for Opus 5.5. This is the last task and probably the most interesting one, building a 3D fighting game using real 3D characters. For this test, I used Mixamo to get the characters and animations. Here you can see the Mixamo links I provided in the prompt. The first link takes us to the character section where we can find a
11:51 lot of different 3D characters and assets. There are multiple pages with around 48 characters per page. The second link takes us to the animation section where we can find animations for things like running, fighting, jumping, blocking, and many other actions. For example, I can select a character and then choose an animation like this body block animation. The character immediately previews that animation. So, this is how I wanted Claude to find the right characters and animations for the game.
12:21 I asked Claude to use the Claude Chrome tool to access Mixamo, find two visually different characters, one hero and one villain, download the required character and animation assets, and use them to build the game. And this is the game output I got. When the website loads for the first time, it loads 38 assets, including the character and animation assets required for the game. Now, we have two characters, Maria with her own set of skills and Warrok with a completely different set of abilities. We can also select different characters,
12:52 and when I switch between them, they actually perform different idle or character-specific actions. The villain even has a pretty intimidating laugh. We also have a two-player local mode, but for this test I am going with one player against the CPU. I will choose hard difficulty because, obviously, we want to make this harder. And we have a control section here. Honestly, there are a lot of controls, so I cannot remember all of them. I might need to keep this screenshot on my phone. We have the movement keys, space to
13:22 jump, F and J for slashing, G and K for heavy slashes, R and U for attacks, L for kick, and shift or I to block. So, let us start the game. There are also sound effects during the fight. So, if you are watching with headphones, keep the volume at a comfortable level. I am playing as Maria, so let us start round one. I will try to speak less here and just let you watch the gameplay.
15:31 >> [snorts] >> And after completing the fight, we get a proper result screen showing the rounds won, hits landed, best combo, damage dealt, and attacks blocked. So, that is the final test. What do you think about this 3D fighting game? Let me know in the comments. I will be posting more test outputs on X, so make sure to follow me there. That is it for this video. Thanks for watching, and I will see you in the next
16:02 one.