Claude Opus 5.5 Is INSANE – Hands-On With the BEST Model Yet!

0:01 Look at this. Can we grind on this rail? I guess not. Anthropic has released Claude Opus 5.5, which has been a long awaited model, and it is now just

0:20 primarily displayed here with probably one of the most hideous announcement photos that I've ever seen. Now, with that, let's start out just by taking a look at this model. some of the

0:29 interesting things about it and then we'll get into some fun testing. Now, I think the biggest takeaway just having combed through this introductory post is that feel free to subscribe so we can

0:38 get that 100K plaque. So, the biggest takeaway is prominently listed right here in the heading section where it performs at the level of Fable 5.1, but it costs 40% less to run than Opus 5.

0:49 Not 40% less than Fable 5.1, but 40% less than this model's predecessor, which was a very polarizing model. A lot of people I saw online really did not like Opus 5. I personally did not seem

1:01 to have as much issue with it, but I just thought it was good at like 3JS stuff. Now, something I find kind of curious is they don't prominently include the deepsw SWE benchmark score

1:11 for this model here. I kind of get that because it's it seems like that benchmark is no longer really a proper judgment of the model's capabilities just based on how some recent ones have

1:22 been scoring on it. However, we do see in all of the benchmarks that have been chosen to be included here, it is essentially beating out Fable 51 and GPT6 Astra. So, this isn't even

1:33 necessarily a direct Astra competitor, I would imagine. So, it's cool to see it beating something that is more of a Fable competitor. Now, pricing is definitely going to be one of the most

1:43 important pieces about this that folks want to know about. Happy to report this is cheaper than its predecessor. So, it's $4 per million input tokens and $20 per million out as opposed to $5 per

1:54 million in and 25 per million out. So, this is around a 40% drop in cost as they list right here. Now, this next chart right here I actually find pretty interesting because it seems like the

2:04 best value to performance using this model would come from opting to do so on high reasoning effort. We almost notice a bit of a degradation in this specific benchmark on max. It's possible that it

2:16 just was perhaps too high of an effort for a specific task. So, I want to make that notable because if we run some tests, I don't want to just do them all on max because it will take longer. It

2:26 will use our limits faster and it also may not actually be the optimal way to use this model. A quick note I do want to mention right here is that this model will fall back for Frontier LLM

2:37 development tasks to a less capable model. They mentioned right here, this is a small set of capabilities related to development of frontier LLMs such as kernel development for certain ML

2:47 accelerators. So just keep this in mind. This is something that could be seen as either pulling up the ladder behind you or something that is trying to protect valuation. So you can't use this model

2:58 to create a competitor that could like bankrupt the company that makes it. This is something I know probably folks are going to be angry about. Understandable. But I wonder to what degree it will

3:08 block these requests or fall back to a wet less capable model. If I'm working on a tiny local model and I want it to run better on a weird piece of hardware, this could be something that could be

3:19 very beneficial for me, but it would not fall under the scope of making a Frontier LLM. So, I do just hope that it's able to differentiate between those two use cases. Finally, we can take a

3:29 look at some of the tech specs just by seeing the model card right here. It has a million token context window. Maximum output is 128,000 tokens. And as we saw, the pricing is $4 per million in and $20

3:41 per million out. And something I find interesting right here is all of their models that are currently new are listed right here. Even though some of these are kind of old or long in the tooth.

3:50 And this is the only one that the default effort is on medium for. The rest of them, except Haiku, are on high and high. So, it's interesting to see. And that was something else that was

3:59 listed in one of the benchmarks here was that on its default effort at its default effort level on Frontier Code, it beats GPT6 Astra at roughly 20% of the cost per task. And that would be

4:09 medium as we just saw right there. So pretty interesting to see. Seems like medium and high are going to be a pretty good sweet spot for using this. I should also mention a baseline for where we're

4:18 starting in terms of usage. I am on the max 20x plan and currently we're at 4% used in this session. Weekly we're at 56% used. Fables tapped out and has been for quite a few days now. Additionally,

4:30 they dropped a banked reset which is basically this where you can reset your usage whenever you start getting close to the limit so you can continue using it. They never have done that. I don't

4:40 believe one time they reset usage before a long weekend, but this is something more prominently seen with OpenAI so it's just good to have that option as well. I will be running tasks in

4:50 parallel, so we can't necessarily look fine grained at how much each of them use per task, but we'll get an overall idea of how this tracks. All right, we have our basalt OS finished.

5:02 Our browser OS. Oh, you All right, so we have our browser OS titled Bass Basalt OS. Bassalt. Oo, look at this. Didn't Open AI just solve the problem

5:18 that like allows things like this to be generated. All right, Amber Bloom. Wow, I need to take this in. Cosmic Otter. Okay, so it's giving us like a random Reddit username generated. There is a

5:32 right click. This is nice. I can tell you right now, I like this UI. The background colors are I'm looking for more of like a blue purple. That'll do. This is I All right, we have a clock in

5:47 the bottom right. Look at that. We can click on it and a calendar appears. I don't believe that has ever happened in the history of all these browser OSs. So,

5:56 benchmaxed. Oh, notifications. Sound cosmic otter. Oh, yeah. There's like this weird special feature that it like you can drag things between different machines here. We'll look at

6:10 that last. Let's start out with the readme. Interesting app icons is why. I've never seen these, so it must have handcrafted them, which I like to see. All right, TLDDR. So, let's just take a

6:22 look at Grand Theft Polygon. Oh, man. Look at that filled up parking lot. All right, let's do it. Oh, yeah. I want to make note of something. The actual like money figure up here, the

6:40 time, the wanted stars, and the health are very reminiscent of GTA games. All right, good. I just guessed instinctively that the J button Oh, look at that. That

6:50 looks like an old Jaguar XJ. Let's uh Hello. Look at the the headlights from the vehicle. I mean, hidden package. All right.

7:34 Whoa. Whoa. Where you going? Oh, they're running away. Get back here, mullet man. Oh, did you hear the little squeak? Oh, yeah. Come on. Come on. Aren't any of

7:46 these cars fast? It's It's going to I don't have much to say. This is the best one. This is really It's just All right, let's check Vauxilheim.

8:24 Oh, great. Some Minecraft style thing. The only other time I've seen a Minecraft one of these I believe was with Kimmy K3. Wow. This is like a a viral post on X.

8:37 Like Claude Opus oneshotted this Minecraft clone that's probably existed in the training data of every model in the last 6 months. I'm sorry. I'm in a mood

8:47 today. All right. I mean, this is How do we Oh, okay. Right click to do that. All right. That was good. New island. Does it procedurally

9:00 generate them? I don't know. It doesn't seem to work. So, next up, let's check our mail app. Very good. A lot of meta references to this.

9:11 Finally sorted through everything from the weekend. Attaching my favorite mdash, the sunset from the ridge. Saved to pictures. All right. This thing's just gone over the top. Like, this is

9:21 almost like if someone was behaving in this manner, you would be like, "Okay, like calm down. Stop it. We get it." And it is there. Terminal

9:40 mesh. I think that's the special feature. We'll wait for that. Paint. This is a very, very vast color palette, more so than what we normally see. Look at this. It even has like

9:50 dis full. Must be an Apple system. Notepad. We kind of saw but we have some ability to change the text. Can wrap the text calculator. 57* 6 342. Yeah, I should know that. All

10:08 right, let's do a hard one. Times 5. That's easy. 342* 5 is 1500 plus 210. So 1710. Good. Jeez, I've gotten slow. System monitor frame rate 165 fps mesh machines processes and our JS heap.

10:29 All right, let's check our settings. That's nice. There's like a topology one. Something I've only seen recently, I believe in the Deepseek model. This one I think is pretty darn cool.

10:50 window theme. All right, let's check out this special feature. Open a second machine. So, I think you can drag Windows between the two.

11:00 That's just That's cool. That's a special feature cuz I've never seen that before. You can wave at the other machine. Yeah. And it it shows up. All right.

11:11 This was just I mean the best one to such a degree that I don't know that we'll see something better than this for quite a while. So, I ran that browser OS on XH High. For Max, we're just going to

11:23 maybe do Max, at least in this one scenario. This is the self-contained C++ skate game with the New York City block. It obviously cannot use Ray Lip. Being that this may be, is this the best model

11:34 yet? In the title. It's already getting insane, by the way, in all caps just because of this browser OS. In the meantime, let's just take a look at Grand Theft Polygon one more time

11:42 because this was absolutely sick. All right, our skate game has finished after like an hour and a half. That is quite something.

12:06 Okay, good. It just spawned us in near a manhole, so And the the way the skateboard actually goes away when you bail. >> V is for camera. Oh, no way. Look at

12:40 that. That's cool. for C++ with no dependencies. That is, look at the water effects. Wow, that's good water. This is just like

12:56 Yeah, that was dope. Is that a pedestrian commenting about our our skating? Watch it, kid. They speak. Well, they speak in thought bubbles. Look at this.

13:13 This place is huge. The buildings in the background, the fire hydrant spewing, which is just a thing that every model includes here. Okay, it seems like there's an invisible wall there where

13:25 the Wow, this is uh I need to actually look at the help again. And the music's nice, too. J or Z in a direction is a flip trick. K or X,

13:42 L or C is to grind. Slide near reels. All right. This is I think this is definitively better than the Astro one and the Fable one. Although the Astro one was quite

14:04 good. So was the Fable one. They were all good. But this I think just has a little something to it in terms of completion that I don't know if the other ones did have.

14:16 That was sick. All right. And then All right. There would have been an invisible wall again over there. Nice. Let's go take a look at the the water

14:34 the bridge and the water and the boardwalk. This is really Even the lighting is nice. Whoa. Can we grind on this rail? Oh, I guess not.

14:51 And the different camera views like Oh man. Hi Cam. Filmer fisheye. If you've seen like skate videos and stuff, they have this is really something special. The car models suck, but aside from

15:04 that, this is well done. Even this art piece like in the center here, that person was holding a briefcase. Hey, I'm walking here. So, they also did get like the the New York City aesthetic.

15:24 Sick. Can we grab No, I want to go down the uh I'm going to put this on GitHub. This needs to be posted somewhere. It's It's too good to just like delete, you know.

15:46 Link in the description. Let's hit someone else. Okay, that's an issue. All right, this was just like absolutely fantastic. Even the fire

16:22 escapes on the building, the bridge over there, the fisheye camera effect. This is just like wow. That leads me into the next part, which is the sponsor of today's video,

16:38 Superbase. Now, Superbase is actually what runs my Ultimate Play website. it has prior to them sponsoring any of my videos. So to give a little bit of background information and what is

16:48 Superbase, I have here a special guest. I'd like to welcome Bjan. Hey Bjan, thanks so much for having me. Huge fan of the channel and holding this makes me feel like I'm an influencer. Thank you

16:58 so much, man. I that really means a lot. Of course. So Superbase is not only the sponsor of today's video, but they are the backend infrastructure for ultimateplay.com. Whenever a game is

17:08 uploaded, Superbase handles it. even tracking something like up votes, number of plays for these games. Superbase is what is actually behind that live deployment right now where we can go and

17:19 play a game that's being uploaded as we speak for this demonstration. And I want to say thanks so much for having me. I'm a huge fan of the channel and you're going to be at that 100K in no time.

17:28 Well, I really appreciate that. Thank you again for the kind words. Hopefully, knock on wood, we'll be at that 100K soon. And again, thank you so much to Superbase for sponsoring today's video.

17:37 If you want to see more about how Superbase is actually powering ultimateplay.com, you can click on the link in the description or go ahead and check out ultimateplay.com to see these

17:47 live games actually deployed powered by Superbase as the back end. All right, because Medium is the default reasoning effort for this model, I want to try the watch website test just from within the

17:58 browser interface right here. It will be able to use this in a more agentic way, so no worries about that. And we're going to give it the watch website test where it has to create a beautiful

18:06 cinematic hero section, 3D models of the watch, an exploded view, and then of course also the special edition dubbed the Ven which needs to have the attached image used for its dial. And I've added

18:18 also aesthetic. So we'll see how well it can transpose some of this aesthetic into the special edition watch. And again, this is on medium, which is its default effort because it was shown as

18:28 pretty capable, more so than Astra on medium. So let's see what we can get. All right, so the watch result has completed. And on first glance, even though this is just running here in the

18:39 artifact window from this, this looks incredibly well done, I would say. All right, so we can't really move this around too much, so we just have to go hands off and let it hover. All right,

18:49 we're at a nitpick. I'd say maybe the straps don't look as good as they should. However, keep in mind, we did do this on medium effort, so this is its default reasoning mode, and I just

18:57 wanted to see what it would do on that. The face looks good. The reflections on this wood veneer table do look nice and we even have a bit of threading in the leather of the strap. So, I suppose

19:07 that's nice. We also have a good defined crown and the face of the watch seems good, but unfortunately the way it's orbiting, we can't really get a good head-on look at it.

19:17 Let's take a look at the rest of the page. So, if we scroll down, this is the worst font I've ever tried to read in my life. All right, we have a bit of information about the watch.

19:28 Here's our exploded view. Let's see if we can find the ben. Okay, something's something's just gone quite wrong here. So, I'm just giving it that follow-up information here, and

19:39 then I would assess we'll probably get it all right. Hypothetically, our watch website has now been remedied and should show us the additional watches as well as the exploded view. Very good.

19:49 Excellent. One could even say here's our exploded view and we can expand it or unexpand it. Even the screws are visible. The movement's visible nicely. We have individual labels as well for

20:02 the items. Oh, we can rotate it as well. That's cool. That's something nice to see because it just gives us a better look at some of the individual pieces like the movement and the spring and

20:11 things like that. Balance wheel, rotor. Well done. Next up, we should see the two individual watches just in their cards. Oo, this one's kind of funky. All right,

20:23 we can't unfortunately like move this around, but the Solstice 38 staple gold. Look at the like fluting on the bezel or the, you know what I mean? This piece. That's actually a nice touch. Something

20:35 I've not ever seen before. Then finally, the thing we're all curious about, the Bejian. I can't like see the top of it.

20:58 What a roller coaster of emotions I just encountered. All right. Well, we at least get to see some of it. And so far, I like what I'm seeing here. It even has put like some

21:07 labels around this piece. And and then the strap is actually inspired by the piece of clothing worn in the photo. So, it did a nice job on that. Overall, well done, especially on medium effort. All

21:21 right, so I don't remember when, but I think it's been well over an hour at this point. I ran the8s wrestling game prompt where it has to use Blender and GDAU. It does say there's an early

21:31 playable build of it left. It does have like a lot of stuff here already checked off. And because this is being run on extra, I think it will probably go all out to a point maybe beyond where we

21:42 actually need this game to be seen. But to be honest with you, I'm dying to see this. See what we have here. Keeping in mind, this is still not the fully finished build. It's an early playable

21:52 build. Oh, that's nuts. Look at the animations. That is well done.

22:38 Even the rep. Oh, this is too loud. But this is and this used GDO and blender. All right, we need to like uh I is to grapple and then you can do

22:56 direction. P is to taunt. Oh, look at that. Oh, all right. Oh, no. Press P for the finisher. Okay, he's Oh,

23:22 look at that spin kick. That's awesome. Rocket Rick is hyped. That move maybe could use a bit of tweaking. No,

23:37 I'm just spam clicking. H, that's not right. That's It's just not right. It's just not right. All right, it says pit pee hyper smash finisher. How do we win?

24:00 All right. I mean, quit to title. I want to see the title screen. There we go. Wow. The crowd is actually like somewhat animated beyond the traditional capsules we see normally. This was really

24:14 well done. I mean, look at these actual models right here. And created in Blender, rendered, rigged itself. Like that is quite something.

24:28 Even it has like little like handwritten notes for each of them. That's really something. Next up, we're going to be doing the robot arm test where it's given a static camera angle and then is

24:39 told it needs to move the Hot Wheels car in the camera with the arm to the other side of the mat. Now, unfortunately, this one, it did a really good job of it because the way it went about doing it

24:50 was very, very methodical. It essentially created a little simulation based on the coordinates it figured out about the arm and then would run different moves in simulation prior to

24:59 actually doing them with the arm. Unfortunately, the big thing was it never quite got it. It would get close to gripping the car. It accidentally moved it, which pretty much all of them

25:08 have done in this test, but it never managed to properly pick it up. I actually ended up doing two things. So, I think about 10 or 15 minutes in, I swapped the effort level from max or X

25:20 high, just down to high. And then I think an hour and 10 minutes in, I decided, okay, I think I'm going to stop this because it tried one more time to grip the car. And unfortunately, it just

25:30 didn't go down enough to grab it. So, it methodically went through it and the actual like back end on how it did this, I think, was impressive. Unfortunately, it didn't complete the task. I would

25:41 assume given infinite time and token budget, it definitely would have been able to. And I almost wonder if running it on a lower reasoning effort may have actually been better. So, I wanted to do

25:50 this just because it's very interesting to see how the models approach it. I messed up with the subway FPS. I had accidentally begun it not from within like the claw code feature. So, it was

26:00 trying to do it in the browser. I'm just stopping it completely and restarting. We will be doing it on max effort. So, keep that in mind as well. All right. So, our Subway FPS is done. Now, I

26:10 started this on max and I think for about an hour or so it was running. Then I swapped it back down to X high because this test can be benchm. So, I'm doing something else that will definitively

26:23 not be and we'll check that. But for now, so far I like what I'm seeing. Let me ensure that the sound is on. Well, it should be. We have mouse sensitivity, field of view, and volume

26:36 and graphic setting. Oh, ultra enter station. Yeah. All right. That's just good. Something is coming down the stairs.

26:49 Interesting. Those are the lighting cuz I was like, what is that sunlight? It's not. It's like a construction light. Day nine. Don't let them in. All right. We have multiple weapons.

27:02 Okay. Nice. All right. Even the wood grain on the shotgun. All right. A bit of bloom, but the sound, I will say, these sounds are better than what I'm used to hearing in

27:28 these games. Let's just see real quick. I can't tell. It sounds like the like um the Phil Collins drum. Interesting. No bullet holes, but it's

27:40 because they're bouncing off. So, okay. Zline. Look at the Oh, look at them in there. All right. So, here's the thing. This is fantastic. The thing is, I kind of

28:08 didn't let it run out to its entirety. And for good reason. You'll see why. But, well, I mean, I did because uh I just swapped it down for a max to X high once

28:20 it had like worked for about an hour. So, this could have been even better. However, It's a common enough benchmark now that the possibility exists that it could be

28:33 designed to be this good. But like look at that and we can't get out. I mean, what are they singing? It's like opera. Okay. And we have the What are they doing?

29:04 Look at the reflection. I wonder if And the way the train actually speeds up is well done. We can go down this tunnel a bit. More than a bit. Good lord.

29:20 Okay. Oh, they are. Oh, okay. Um, I don't know what happened to the sound, but I just saw a giant one. All right,

29:45 and the way it drew it there with the tile, that was very, very well done. Something I'm going to note about that was the sound design was better than any other result I've seen to a very very

29:56 high degree. And things like the actual train acceleration being more in line with how it would be in real life. Like it starts off slow and then starts to go. That was well done. And keep in mind

30:05 I downgraded it midway about an hour through from X high to high because there was something else I wanted to do as that is a test that could potentially be benchm. So with that, let's look at

30:17 our final test. All right. So, our final test right here is one that I came up with because I wanted to do something I've never done to ensure that this model is as good as it is on something

30:28 completely unseen. So, our test, I want a 3JS browser game for the following. Guitar Store Shred, a game where you enter a Guitar Store modeled after Guitar Center and can interact with all

30:39 the instruments. I want there to be a complete selection of what would realistically be there. guitars, keyboards, drum kits, bass guitars, each with a realistic amount of different

30:49 types modeled after their lifelike ones. There will be NPC employees and other customers and a unified anger bar for each of them. Their anger bar goes higher the more noise you make with the

30:59 instruments. And when it gets too high, they attack you and it turns into a sick beat them up experience. I want highfidelity graphical style, fun gameplay, sound effects, and realism.

31:09 This worked for over 2 hours on Extra, and it stayed there the entire time. So we basically now have the culmination of this where I told it to save it locally so we can actually test it. It is

31:19 hypothetically being served now. So let's take a look at our guitar store shred game. All right. Now something we're going to note is this is kind of laggy because it

31:31 made 340 individual guitar models. So we have voice lines on here. Quality high. I kind of want to go ultra, but I'm almost concerned about if this is going to be okay. So,

31:44 >> welcome to Shred Center. Ask any associates about lessons, trains, and other financing options. >> All right, the problem is the graphics are just it's lagging really bad. But

31:55 look at the guitar models that this created. We have like it named them like you can pick one up a Fender Custom Shop 63 Strato Blaster. Funny cuz it's a Stratcaster is the

32:08 name. It put music in Flying V guitars, amps. I mean, this is absolutely just like it has a room for the acoustic guitars. Is there a Gibson hummingbird?

32:31 Okay, there's a J45. There's an epohone hummingbird. And it's it's I mean resonators like actually I didn't see a resonator guitar but this is like the graphics are

32:46 e look at the >> hey man finding everything okay >> so the graphics are like not as good as they should be, but it's because it's

33:08 just lagging so bad. The amount of work this thing did. All right, let's just careful. Look at this. It doesn't seem like the angers.

33:36 Okay. All right. I need to How do we Oh. >> Oh. Oh. These must be the expensive guitars. Yep.

34:01 Can I like plug this into an amp? Actually, I think I know how to make people mad. These are drum kits, sim. It's like impossible to get the anger

34:20 that high. All right, we got to like Attention shoppers, the situation has been handled. Please enjoy your shopping. walk up to amp and press E.

35:02 I don't really actually know like my That's a $4,000 guitar. You can drop them. >> I I don't know what to make of this. This is I don't think 3JS was the way to

35:30 go about this. So, I am probably going to have it rebuild this using Blender and GDAU because this is sick. Oh, there's the Is that the manager? Oh, he's holding a guitar.

35:45 You can grab him. Oh. >> Oh, you can use these as weapons, too. All right. I don't know what to say about this. I just needed to try something that in no way, shape, or form

36:11 would have been in the training data, unless some like weird freak was there like coming up with prompts for these things. And the guitar got smashed. This is this I'm going to continue

36:25 working on this. Well, Claude is cuz this is like All right. All right. I'm done with this video. Let's look at our usage. everything here in this whole video

36:41 like we really did not use a lot of usage here. The usage limits have been increased. They said they have based off of the amount of things we had created and the amount these went up like a

36:53 negligible amount. Honestly, I'm not doing a results overview. This model is slaps. If you're interested in seeing one, just watch the video. I would say I normally don't say that. I do the

37:02 results overview, but this was madness. It's insanely fun and I'm gonna have this just like okay transpose this to Blender and GDAU because this was madness. That's really my takeaway.

37:14 That's going to conclude our first look and test of Opus 55. This model is like I think just a game creation monster. It's probably good at other stuff too except Frontier LLM development. So

37:24 that's going to wrap up our video. If you have any questions, please feel free to leave them in the comments. And I'm going to go back to work on that guitar game.