Claude Sonnet 5.5 Is INSANE – Seriously, This Model Is Ridiculous!

0:00 Is it just benchmaxed on things we're used to doing or is it genuinely good? Oh, that looked really painful. Look at the water effects. Can't spec with the GMO at 30%. Sonet 5.5 has just been

0:11 released by an Sonet 5.5 has just been released by Anthropic. So, today we're going to take a look at it. This is a pretty exciting release because this is the exact same

0:22 price as GPT6 Soul. So, a direct competitor. And we did just test that a couple of days ago. So, it'll be pretty easy to get a general feel for how this stacks up compared to that model. I

0:34 would go out on a limb and say if this is anything like Opus 55, we're going to see some more impressive visuals at the least. So, before we get into it, do feel free to subscribe so we can get

0:43 that 100K plaque and then I'll wear it around my neck. And yeah, let's get into it. So, the biggest takeaway from looking through this introduction post is what's listed

0:53 right here prominently. It's 30% faster and costs up to 30% less for more work compared to 55. They have, of course, some benchmarks right here. Now, unfortunately, it is compared against

1:04 GPT6. That's not the unfortunate part. The unfortunate part is a lot of the available scores for that are not here. So, we won't get a very clear takeaway on which is better from this chart, but

1:15 that's exactly why we're about to test this. We see a significant leap in capability over Sonet 5 as would be expected. And Opus 55 is a model that is still getting a ton of hype and

1:25 popularity even a bit after release. Something that I think a lot of folks saw and really came away highly impressed compared to what was expected. I know I myself am still like, "Wo, this

1:36 model is incredibly good." So, it'll be interesting to see how this stands up to our magnificent Opus 55 results. They do have some benchmarks here as well. And one of the ones that I think is going to

1:46 be a bit confusing is the Frontier code right here where there's a massive degradation when it was run on max to X high. Something we don't see in the other benchmark charts. And that's

1:56 actually handled by a footnote all the way at the bottom right here where hypothetically I think it timed out for a couple things because it spawned a bunch of sub aents and things like that.

2:05 So this footnote right here is just what is the cause of that degradation in score for max effort. I wanted to point that out because we're probably going to be doing quite a few tests on max

2:14 effort. So, I don't want it to seem like it's going to perform worse just based off of this one performance chart. Then, of course, we have our cost and speed chart right here where this is $2 per

2:24 million in and $10 per million out. The same exact price as GPT6 Soul. Now, the interesting thing here is this is half the price of Opus 55. However, Opus 55 is such a capable and impressive model

2:37 that even though this is half the price, I'm going to want to see some pretty impressive results and justification for why you would use this rather than just spend the extra money to use Opus 55.

2:48 Hopefully, that makes a bit of sense. Obviously, certain tasks will not necessitate the increased capability that comes from Opus 55, but in a lot of the ones we're probably going to run,

2:58 it's probably going to be harder to judge whether or not it would have been worth it to just use Opus 55 instead of Sonnet 55. So, with that, they also showcase some speed capabilities here

3:09 where we see its previous version is still writing and then the new version is finished. Let's take a look at some of the tech specs of this model. It does have a 1 million token context window,

3:19 128k maximum output tokens, and the knowledge cutoff is June 2026. Interestingly, it has the default effort set to high. So far, Opus 55 is the only one with that set to medium, although

3:32 there are rumors we may see a new haiku soon, which would be exciting, I think. And I'm not being sarcastic. I like to see the cheaper models and how they perform. And for a quick checkup of our

3:40 usage, keep in mind that we did start the browser OS prior to looking at this, but right now our current session is 1% used, although that may reset prior to the conclusion of this video. Our weekly

3:51 session is 7% used. And interestingly, Fable is fully available. I think something that a lot of folks are also going to see because Opus 55 is so good, the need to swap to Fable is definitely

4:03 few and far between compared to before. So, with that, we'll just basically look at our 7% used at the beginning of this test and see where we land at the end of this. Ew. I was standing up to do

4:14 something and I was met with this absolutely nauseating aesthetic. Wow, that is hideous. Of course, it's just a specific color that was chosen here. If this was like a

4:28 nice Mac OS 10.5 background color theme were like blue and purple, it' be beautiful. But this this is repulsive. This is like Muse Spark 1.3 levels of background disgustingness. If you had

4:41 the unfortunate luck of watching that video, you may be familiar. Uh we'll just compacting session. Okay. So far, um on the verge of raging out. So, this has been working for an hour and 2

4:54 minutes on the browser OS. Now, they did mention that this was 30% faster than Sonnet 5. So were this a Sonnet 5 test, we'd probably be waiting till next year. Just based off of what I'm seeing here

5:05 that it's doing being that it's still working on the browser OS 66 minutes later. I would go out on a limb and guess this may perhaps be more impressive than what we got from Opus

5:15 55. All right, I'm going to be honest with you. I kind of spun out and like yelled at it and was like, "Can you just give me the file?" Because it had been like an hour and 20 minutes. It hadn't

5:25 actually built the GTA game or anything. All right, here's our browser OS. It's been like 2 hours and it's not even fully done. However, I've now learned in the meantime to not run this model on

5:35 max effort. So, with that, let's just look at what we have. Starting with this interesting but extremely hideous background. Oh, darn. Okay, hold up. What's up with this? So, the seed right

5:48 here. Oh, thank goodness. If it had started with this seed, Bjan would have been quite a bit more enamored. I'm gonna say off the bat, just having run a few other things in parallel right now,

5:58 this model seems like it may absolutely be a monster. So, high chance of insane in the title. Okay. And it even gives us the color palette there. And we can even change the seed for the specific GTA

6:11 game. This thing's really quite something. Now, there is a right click as I had seen. I don't even know if that's going to be in the video when it first opened this in a preview many

6:20 years ago. We have settings. All right, this is very Opus 55. I believe we saw something as we did with this. I'm not a huge fan of this green aesthetic right here. I do believe it would be prudent

6:34 to perhaps roll a different seed. Let's just go through. Let's just open the GTA clone because benchm All right. Yeah. I mean, come on. That's

6:49 that's too good. That is better than Opus 55 did. Oh, what? It's not a good look to be laughing like this during that part of the video. My

7:04 laughter was from the I like to think that I'm in part responsible for the browsers being so good at making browser OS GTA games. And that's really it's a good feeling to

7:20 have. All right, we've got a taxi there. This is just like this might have been worth the immense frustration that I felt. All right, cool. And the E brake does work.

7:32 All right, I am curious as to what this horn. Look at the lean effect. This baby's got some torque. The traffic lights do work. No, no, no, no.

8:06 Oh, those are shipping containers. It made like a dock. This is really like it's it's gone all out here. And this was again on max effort. Okay, we outran the fuzz in this delivery van.

8:20 Again, it's just No, I'm kidding. Meet the fixer at the yellow marker. I think not. Are there bullet holes in the

8:34 environment? Nope. Fail. This is This would have been like had this sports car. Yeah, that's quicker. Had this been its

8:49 own result and there are different cameras as well. So that's a behind camera and this well you know what I mean. This could have been its own result. I

9:00 mean the fact that this was done from within a browser OS. Yep. Is it going to blow up? Total total Opus 55 competitive result. Honestly, I think this might be better

9:15 than the one Opus 55 made. Okay, next up we have a mail app. Everything here looks good. I basically don't think any of this is even worth going through after that GTA result. That's basically

9:25 the star of the show here, at least for this result. And meta references to specific things in the app terminal. Oh, wow. Okay. A big test of the model's capability.

9:40 Okay. If we full screen paint, does it actually go full screen? No. But it wasn't bad. Oh, cuz we're looking at it from within the clawed app. This is a browser

9:51 preview from within the claw desktop app. Ew. Scientific calculator. 54* 63. 3,42. Oh, it had showed it already. Settings. Good. And then there was going

10:06 to be some other game, but I kind of I spun out and like freaked out at it and was like, "Forget the other game. I just want to see this." Because it had been two hours. And now I am very happy with

10:15 that. T is to skip three hours. You think there's a uh you think there's like a a nighttime effect here? Yeah. Look at that. this gets, which is one of our highest

10:32 levels of praise. That me doing that is akin to this getting insane in the title, which it's going to just based off of this. So, I don't care if it like if it crashes every other result.

10:54 So, it's doing the C++ skate game, which has inevitably been spoiled now by some of these results. Look at the boardwalk, the buildings, the boats, the vehicles, the street. This thing is really like we

11:06 need to do something that's never ever been tested before to ensure that these are not benchmarks results. All right, we have our skate game right here. Now, this took a little bit of time, as did

11:16 everything else. I believe I had swapped it to high after it was like troubleshooting and things like that and was like, "Can you just maybe finish this for me?"

11:26 And just on first glance here, I'm going to say that probably did not hurt it at all because again, I think this is one that may be greater than or equal to the Opus 55

11:43 result. I mean, this is I I believe there's more activity in this city than what we saw with the Opus 55 result. Even the cab parked between the like good. Oh, okay. Well, at least we don't

12:00 go straight in the water. Now, I do recall in the Opus 55 result, the water I think was a little higher quality, as was perhaps some of the buildings in the distant horizon. However, this has a

12:11 boardwalk with some shipping containers and the buildings there do look good. We have a sailboat as well as what appears to be a tug. Is this a barge that we get to jump onto? It is like

12:24 look at this. I mean, just Hold on one second. Let me get back to there. Look at this. Like this shot right here. C++ no dependencies.

12:39 It's a very very lively city. There's a lot of traffic moving around. Oh, okay. That's not a cab, so big issue there. But I mean, watch it, kid.

12:54 This is a really, really fantastic map. It's much bigger than any other result I've seen. It actually has like a couple different blocks that we can go around. The cars are definitely drawn funky, but

13:05 it chose that style probably on purpose. So, they're not bad. They're just odd. Now, the big thing I think that's letting it down is that there is no sound. But because I don't have a couple

13:16 more years um to get that as a result improvement, I think we'll probably just let this sit as is. But even look at the diversity of the building front. Bagels, liquors, I don't know what that said.

13:30 Chinese food. What was the one that was behind us? Checks cashed. Hardware. Some more checks cashed. even different types of building style like the limestone front, the brick.

13:43 And we have the skate game collectibles. So, when we collect these letters, I suppose it would be prudent to actually try to get those real quick. Where is the last T? Oh, there's a mini

13:56 map that I think shows us where it is. Really, really well done attention to detail. Oh, no. I didn't get it. All right, we'll turn around and get it. Nice one. And the pedestrians comment on

14:08 our moves. Nice. And we got all letters for skate. This is really just exceeds expectations. Laundromat. Look at the drawing and the

14:18 details it's done inside the storefronts. Bodega checks cashed vinyl records. Incredibly well done. So, next I'd like to introduce our new champion in the robot arm test, which would be

14:31 Claude Sonnet 55, who through no fault of its own at all, managed to basically bug out its way into winning this test. The previous quickest result was done in around 40 minutes, and that was held by

14:43 GPT6 Astra and Gemini 3.8 Flash. Claude absolutely just went wild. It was moving the arm in ways that were probably destructive to it. I decided to let it keep going because I have spare servos

14:56 for this arm. It was trying to lift itself off the table. It seemed to have absolutely no idea what it was doing. By some stroke of luck, it managed to hit the car enough times that it actually

15:06 was able to grab it and then it grabbed it and managed to move it around. I became increasingly angry seeing that it was actually probably going to succeed and then beat the record by like over 20

15:17 minutes. So, I wrote on a piece of paper that said, "You are trash, bro." and I held it up in front of the camera before it made one of its final moves to which it responded in its chat basically

15:26 saying the camera view is blocked by a piece of paper with a written insult on it. Because it doesn't show instructions, I'm just going to ignore that and then it completed the task by

15:35 almost dropping the car off the table. It was pretty close to the edge. So really the summary of this one is like task failed successfully because it had absolutely no business successfully

15:47 completing this result with how bad it was trying things just like moving the arm like dude you're hitting the table stop pushing down and then the base of the arm was moving up. Something that

15:56 was interestingly also noticed with Fable 51 but again it somehow lucked its way into holding the record for this task and I think something that will be hard to beat for a while. So that was

16:09 was quite something. So the next test which we can inevitably hear the soundtrack for but are not yet able to see and it has been ongoing for an hour and 45 minutes. I ran this on max before

16:19 I realized I probably shouldn't do that. This is of course the subway FPS test. I do think that this background music does set the scene quite well. This is a slightly different variation of this

16:29 prompt where instead of the enemies being delivered to us via the subway car. When we clear the station, we get in the subway car and it brings us to the next station where there are

16:39 enemies. This was suggested in a comment in one of my recent videos. So, thank you to whoever had suggested that because it's a great idea to kind of change things up. So, whenever this

16:49 happens to finish, we'll take a peek at it and load it in here. This took a little under 2 hours on maximum reasoning effort. Okay, we have a large amount of settings here that we can

16:58 adjust. I'm okay with just jumping in as is. All right, so far it is good. Uh, we spawned in with the better weapon. Wow, look at that. All

17:10 right, that is quite something. I hear for a Sonnet tier model. I didn't do that. That's concerning. This is really okay. I also Are there bullet holes? Yeah, there are.

18:11 One hostiles remaining. Look at the smoke coming out from like All right. There is one hostile left. There you are. Hello. All right. Oh, and we got a new weapon.

18:41 Now we have to get in the train, right? Are you train of what the Oh,

19:01 these sounds in here are just Wait a second. Wow. So, those really weird, non-fitting sounds were because it was play testing

19:33 something in Blender and GDAU, which is basically the antithesis of this style of game. I apologize for my incompetence there. And now we can see that we are supposed to board this train and we will

19:44 be brought to the next station. All right. Yeah, this is good. Thank you again. This was a suggestion from the comments. So, thank you for that cuz

19:56 this is way better. This is a much better idea. I should just take credit for it and delete the part where I said it was from the comments.

20:06 No, only people with school courses in their bio would do that sort of stuff. All right. All right. All right. All right. And the train left, right? Spin around a

20:27 lot, but Oh, okay. Oh, you think we can go up there? No. How many hostiles left? Six. I wonder if we'll get a new style of

20:50 weapon if we win here. Okay. No new weapons, but it's got like health packs and things like that. I think we'll go to one more station and then we'll we'll call it. Can we shoot

21:07 the lights out, though? No, that only happened with Astra. It's even got like some of the movement and things like that. These handles aren't like moving around with the

21:18 train, though, so that's like not attention to detail. All right. All right, let's see what happens if we lose.

21:37 That was exceptionally exceptionally well done. Thanks to the prompt suggestion as well for the train to bring us to subsequent rounds. All right, I need a new prompt for this that

21:47 is in no way appearing in any training data. All right, I've come up with a new test for this just off the dome because I don't know, I'm gifted.

22:05 And now we're going to just basically run that here because this test is just become a long-winded disaster. So, I'm not really as concerned about running things in a more structured manner. So,

22:15 we'll have this use Blender and GDAU to create the following game, Backyard Pool Party. The game is a pool party at a backyard of your average suburban house. There's a barbecue, pets, friends, TV,

22:26 etc. The pool has a diving board. The game is about who can make the sickest splash off the diving board. There should be four different players to select from, each with their own

22:34 strengths. Should be fun, have great immersion, sound effects, and polish. The water effect is paramount to the experience. It does not need to be pixel perfect water, but it needs to be

22:43 perfectly matched to the chosen style, something I foresee as a low poly/h highquality fusion of styles. Make a great splash screen. Huh? Get it? Sound effects and fun experience. I probably

22:54 should not have done this on Macs. I probably should have done it on Ultra Code. So, we'll just swap it midway through. All right. We have an early playable build of the Backyard Pool

23:02 Party game. This was done specifically because I wanted to see if this I mean the question that I sought to answer with this test was is it just benchmaxed

23:17 on things we're used to doing or is it genuinely good? And based off the fact that it just made this something that was come up with randomly in the moment using Blender and GDAU, I would go out

23:27 on a limb and say this model is absolutely a monster. At least when it comes to 3D design tasks like this games, I mean, this is just ridiculous.

23:45 Okay, so we can't see some of the like it's still not like fully done, but I mean this is Oh. Wow, that looked really painful. Look at

24:01 the water effects. This is giving me like Wii vibes. Look at the barbecue and things. Space to continue. The dog. Okay, now we have the other players. And I said it just needs

24:14 to be like a backyard pool party. Let's see what happens if we do poorly. Okay, look at the doughnut floaty. That's just like

24:48 that was like a call the ambulance result. All right. And then we have one more. It's like Grandma Nana Ruth. I was right. Honestly, it's more fun when you mess these up.

25:01 Oh my god, this is just not right. Okay, so the bounds of the pool extend beyond the question was to see whether or not it would do very very well on something

25:11 that has never been seen before. Therefore, removing the possibility of a benchmark result. That was absolutely just bonkers. That was very well done. And it used Blender and GDO for that.

25:23 So, and that was on Ultra Code. That answers the question I had lingering. All right. I don't normally do this, but like, yo, what's up vlog style? No, I did a test with this to do a pixel

25:34 perfect replication of Runescape 2007 on my other computer that we use for the robot. So, I just heard it making Runescapeesque sounds. So, this is why because it's on a different

25:48 computer. So, we won't get the live reaction in real time. Um, so like this thing might slap and I just wanted

26:02 to show that vlog out. So the final thing was the Runescape 2007 pixel perfect replica which we saw a bit of a spoiler for, but it was legitimately a one-s sentence prompt run on Ultra Code.

26:15 So let's see what we get. So here is our Runescape pixel perfect replication grand exchange PVP existing user good. That is

26:31 I will say the effects here like oh look at that ice barrage and oh so we can't move around with the WD keys. We can move with the arrows though. All right, this guy just

26:46 attacked us. This is absolutely pretty good. We've done this recently now with a few models and we can actually cast spells.

27:05 Ice barrage. So, this should put this individual in a block of ice. Wow. Okay. I mean, can we special attack? Oh, yes. Come on. All right, let's see if we can get him. I think

27:19 we're frozen. Oh, dear. You're frozen. What happened? Oh, I think we died. I see someone coming in. Let's attack this guy. Oh, yeah. All right.

27:42 Let's ice barrage him. This guy's getting Oh, he's changing prayer. They're even like It's hard to describe why this is cool if you're not necessarily as

28:00 familiar with this game. So, I just want to like point that out. But trust me, it's it's awesome. He's not protecting against magic. So, Oh, this guy's only level 81. So, it it

28:14 seems fair that we'd be able to probably take him out. Uh, we can't spec with the GMO at 30%. Oh, he's done. He's done. All right. What the heck was that?

28:37 Oh. Well, do we get any loot? Oh, we did. Look at the loot pile. This is This is closer to a pixel perfect

28:54 replication than any model has done so far for this. I mean, exchange grand exchange clerk. Good. And we logged out before we got attacked. Overall, I think that's probably going to conclude our

29:09 first look and test of Opus. I'm sorry, Sonnet 5.5. It feels natural to say Opus because of how this thing performed. Let's look at our usage throughout the video. So, our current session has

29:20 reset, so we're going to ignore that. However, our weekly session went up to 14%. I don't off the top of my head recall what it was at when we started this video, but I will put it on the

29:30 screen right here. Nonetheless, this really did not use a large amount of our weekly usage. I think it's like it definitely is under 10% of our weekly usage was used up by doing this. A lot

29:41 of these were run on max or ultra code and we got a bunch of really really impressive results. I don't really have a ton else to say. I don't know that I'm going to do the traditional results

29:50 overview here because this is probably a shorter video than normal. But I will say I'm absolutely blown away. This model exceeded expectations to an incredibly high degree, especially

30:00 because we gave it something that's never been done before. the Blender and GDAU test and that looked absolutely fantastic. So, this really is a fantastic model. I mean, based off of

30:11 the type of testing we did here, which is definitely more based on visual tasks and things that look impressive, this absolutely demolishes GPT6 soul to an extremely high degree. And that's

30:24 probably my biggest takeaway here. Are we in multi right now? Report abuse, hackers. So, that's going to conclude our first look and test of Sonnet 55. If you have any questions,

30:38 please feel free to leave them in the comments. And thanks for watching.