Claude Sonnet 5 Is HERE – Hands-On With Anthropic’s NEW Model!

What happens if we just don't deploy the parachute? So, oh, it automatically deploys for us. What an anthropic thing to do. Today, we're going to be looking at a very exciting new release from Anthropic, which is the Claw desktop app for Linux. It is now in beta. And as of yesterday, I was complaining that this didn't exist. Okay. Also, they did drop Sonnet 5, but I am genuinely more excited for the Linux support. So, with that, let's take a look at some of the interesting things about Sonnet 5. In the meantime, please feel free to subscribe if you want so I can get that 100K plaque. Now, I have combed through this introductory blog post and I do have to say it seems very realistic about the capabilities of this model. So, for those not as familiar with the anthropic model lineup, Sonnet is kind of like a mid to lower tier model. It goes haiku at the bottom, sonnet above that, opus above that, and up until very recently, Opus was the top, but now we have the Mythos class models, which are Fable and Mythos, though currently unavailable. So, for the time being, this sits kind of in the middle ground. It does seem to perform fairly well, but we see compared to Opus 4.8 date here in terms of benchmarks. It is still not necessarily up to the state-of-the-art in terms of coding capability, but its agent coding does seem to be rather good, especially considering the gap between this and Opus 4.8 is not too large. Additionally to that, they do also mention something that I want to prominently talk about, and that is one, the pricing. Though with the caveat that this pricing is an introductory price that is good till August 31st. So from now till August 31st, it is $2 per million input tokens and $10 per million output tokens. Following August 31st, it is $3 per million in and 15 per million out. Now something interesting about that that is in the bottom here at the footnotes. I want to specifically talk about this. So sonnet is an upgrade, but it changes how the model processes text. So it uses an updated tokenizer. The trade-off is that the same input can map to more tokens. roughly 1 to 1.35 times depending on the content type. The introductory pricing is set so that the transition to Sonnet 5 is roughly costneutral at least till August 31st. Now I am personally curious in seeing how this Sonnet 5 model will stack up to perhaps some of the previous Opus models. So I've just pulled up the benchmarks that were shown in the introduction for Opus 4.7 which was all the way back in April. So many many months ago. And we see right here that it got 64.3% on this Sweetbench Pro and Sonnet 5 got 63.2. And in terminal Bench 2.1 for agent decoding, Sonnet 5 got 80.4% and Opus 4.7 got 69.4%. So we can see it will probably be around on par with that if not a little better in some areas which could be good. And I want to see comparisons of this to previous Opus models. I understand that Sonnet has its own place in terms of models being used, but I think we're all interested in max performance at this point. They do also have a couple of benchmarks right here that do include the previous version, which was Sonnet 4.6. And the Leap did go from 4.6 to 5. So, you didn't miss like anything between those two. And it just shows here. I like seeing previous results from the previous version to compare leaps in capability. And we can see here on X high mode, at least for this specific benchmark, does seem quite performant, almost matching high on Opus. So again, seems like it could potentially be a pretty decent option for a cheaper but performant model if you don't want to use Opus all the time in terms of cost mindfulness. Something else I'd like to point out is I've purposely selected Sonnet 4.6 right here in the web chat interface. And we can see that there are four selectable effort levels for this model. If we select Sonnet 5 now, there is an additional one here. So, just to make note, it does seem that there was also an newly introduced additional effort level for this model. I also quickly want to showcase my current usage limits so that we can get a read on these once all of the tests have concluded running. The current session resets in 2 and 1/2 hours. I don't know that the tests will be finished by then, but for our weekly limits, we are at 8% for all models and 2% for the current session. Though, I think that may roll over. So, I'm more focused on the all model usage. All right. So, I'm going to use this on max. I could also put it on ultra code. Perhaps we'll try that at least for one of the tests, but everything else we'll just run this on max effort. Sonnet 5 is also selected here. And if we notice any oddities, this does still seem to be in beta, this Linux support desktop app. So, just keep that in mind. And with that, we're of course going to start with our triedand-true browser OS v2.5. All right. So, I have set it to bypass permissions. And now we'll get started and see. This is using like clawed code so keep that in mind. It's a guey interface for this but we are for all intents and purposes running this in an agentic coding manner. Just to make that clear. All right. So after quite a long time I believe this was probably like north of 40 minutes. We have our browser OS. Okay. I'm going to say significantly overd delivered in terms of the number of apps. Very good. Other than that it's quite basic aesthetically at least on first glance. Let's just go through the big checklist starting with. Okay, there is a right click. Good. We also do have where my head will be in the video. So, I will have to move this what I would assume to be the special feature which is the Echo voice assistant. So, we'll look at that. But let's just quickly go through these one by one down the line. Okay, everything looks good and standard. Okay, that's all right. Right as I was saying that. Okay. So, we going to notice a bit of oddity right here with the um regardless. Good. And that brings us into our notes app. Let's create a new note. Hello. Will it save it as a text file? Okay, I suppose it should. And deleting works, so that's cool to see. We can make new folders and files as well. Next up is terminal. All right, let's do help joke. Why was the JavaScript developer sad? Because they didn't know how to express themselves. Wow. Matrix. Sick. Very cool. I like to see that. It's included far more often than one would expect it to be. Now, the thing I'm noticing here is I can't actually get out of this. So, can I uh All right, let's just Oh, okay. Good. It did end after 15 seconds. All right. So, you're basically forced to just not use the All right, we have paint now. Something that generally goes wrong with paint is the resizing. And this is no different, but that's okay. Painting.png. Notice the care that it took to actually name the specific thing we made. Painting. So, it makes the user feel good, which is definitely nice. We have a calculator. I feel like these haven't been included in the recent ones, which is kind of weird. Let's do 75* 6 450. Good. Music player. I can't imagine it will actually have something that it plays, but I do have a speaker hooked up just in case. All right, that's actually not bad. It even gives us the BPM there. So, I wasn't expecting it to be functional. I thought it would just kind of have those as placeholders. I'm happy to see that we have some gradients here that it was seemingly going pretty hard on golden hour. Oh, yeah. Deep space. Choose file. Very good. And that is a very cool background image being that I did make it. So, all right. About And we have sound settings as well. Controls the music player output volume. That's not bad. All right. the GTA clone time. Crime City 3D. Now, that is good. This is one of the few to actually be at the level of the Quen 3.627B. All right, there are mesh colliders. Now, something that consistently is a failure point with these, let me full screen this, is the police usually when they do catch up to us, they'll just go like right in the position of our car. All right, somewhat, but not horrifically. So, we have trees. There are mesh colliders and we got a notification that we crashed into traffic. This isn't actually half bad. I don't think there's a way to get out of the car though, which would be something I'd like to see. I'm just looking for more pedestrians. I just I had it doing the C++ skate game and I forgot because it's quite slow and and that popped up making me quite confused. All right, let's just get a couple more pedestrians. Nice. And then we'll uh All right. But I don't see a way to get out of the car, but that's okay. It's It's acceptable for a Sonnet class model. Next up, we have Zombie Siege 3D. This seems like it may be cool because when I was watching its Chain of Thought, it was talking a lot about click to enable mouse lock. And it actually has a button for that, which is curious as most times you would just click in the window. But okay, was that a muzzle flash? must have been. There is no sound here, but that's all right. This is actually probably the best second game in the browser OS test that I've received to date. It's 3D. It's zombie style, which I didn't tell it to do. I'm okay with this. That's not bad. Then finally, we have our special feature, which is the Echo Assistant. I'm Echo, your voice control assistant built into this OS. Try open terminal or click the mic and just talk. It's been kind of a rough day, you know. I'm not sure how to do that yet. Try open terminal. Change wallpaper. It's a great assistant. >> All right. Change wallpaper to ocean. >> Wallpaper set to ocean. >> Very good. Help. >> Available commands. >> Good. >> Help. LS. CD less than dur greater than cat less than file greater than. PWD echo less than text greater than. Clear date. Wom. I'm not sure how to do that yet. Try open terminal or change wallpaper to ocean. >> Why does it say Alexa there? You know, it doesn't matter. Joke. >> I would tell you a UDP joke, but you might not get it. >> Matrix >> toggled Matrix rain. >> A that ruins it. I don't want to hear that back. That make that ruins the coolness of it. And then that's going for 15 seconds. So, okay. Overall, not bad. Definitely a feature-p packed browser OS and the voice assistant was nice that it did actually speak back to us. All right, so in 19 minutes and 25 seconds, our skate game did work. This was run with ultra code in clawed code, so the highest expense and effort level. Okay. Oh, okay. Good. Spawning us into the stairs there really did kind of let it down. It's I actually think we're getting some lean effects when we do tricks. Now, there are some like basically the big problem is this looks like a desert skate simulator instead of a boardwalk skate simulator. But aside from that, the player model is relatively like, let's see, Z is for a kickflip. Good. Okay, so it did kick flip the board and not the board with the player, which sometimes like the less performant models do. That's an interesting NPC. The walking animations are not bad. The scale of the player is much smaller than the NPCs, which happens in this test more than one would expect it to. Okay. Can we keep going though? Oh, that was painful. All right. And then that's kind of a problem is we just go into the abyss. What other tricks do we have? X is for a heel flip. All right. Okay. C is for a shove it. All right. F to toggle fog. It is somewhat visible. If you look at the palm tree, you can see when the fog is on versus off. The in the fountain is nice and the stores are eh. This isn't that good. It's not that bad, but it's not that good. So, make of that what you will. Acceptable though. And I wanted to do this. This was done with Ultra Code, though. And the Opus 4.8 would have done a much more intricate job as expected. But I know a lot of folks, myself included, were specifically interested in how this model would compare to Opus, the 4.8 version. So interesting nonetheless. So I had also run the 3D subway station scene here, but Claude reached its max length for the message. It can use excessive tokens. Consider using a lower effort setting. I'm just going to click continue and we'll see. Okay, so we got our subway station result here. It just showed it in line after like a bunch of really messy looking stuff, but unexpectedly it was actually there. So, let's just take a peek at it. Detailed subway station. Oh, it does this because this is the way it's drawing it in the artifact, which is really quite frustrating. So, it screws it up when we try to look at it ourselves. Maybe we'll just peek at it from here. Even though that's because this doesn't actually look half bad. Maplewood Junction. This is clean. The tiles are nicely spread out. There is a poster for the city gallery. Sometimes the text gets inverted more often than one would expect it would in some of these results. Okay, we have going up the staircase there. Good texture on the back wall at least. There are mesh colliders on the walls, okay, not on the pillars. The do not cross line is good. The track is done all right. The benches are there. This is overall it's satisfactory except for some of that stuff but very simple. Okay. And the the clockwatch face is a bit odd. We do have some conduit going on the ceiling. So this is it's acceptable. I'm going to say though the random brass rivets right there are something I don't really often see. So it's good to have that. So in the meantime I had also run the follow-up subway station test which is to turn it into an FPS with zombieoid enemies weapon recoil sound effects. etc. Okay, it's named it last stop. Maplewood Junction. Line service suspended. Track conditions hazardous beyond this platform. Multiple hostile entities confirmed on site. Remain on platform and defend your position. All right, let's take a peek. So far, not bad. Okay, we have footsteps as well. Let me turn my speaker up. All right. Oh, yeah. I would assume these are the hostiles. All right, now the big one. Are there no bullet holes in the environment to be expected? Not like included with the sonnet one, but okay. Not bad. Clean. It's competently done. There's nothing here that wows me, but it is competent. Shift is to run. Let's see what happens if we lose. Oh, okay. Last stop reach service terminated. And we have some stats. Okay, I'll take it. And just something that I always enjoy doing as a follow-up to the subway station scene. Next up, I'm going to be giving this a new test that was introduced in the GLM52 versus Opus 4.8 video. So, we will also have both of those results to do a quick comparison of the quality of this Sonic 5 result. This is on max thinking still, so I would expect this to take quite a while. And the task here is to create a 3D skydiving simulator. This isn't deliberately an open-ended task that I wouldn't imagine that text garbling will affect it. The exact implementation of the simulator is up to you. However, there must be emphasis on highfidelity visuals, smooth gameplay, and the polish that would be expected of an indie game. So, this is a fun one as it's a new test to the channel. What the Okay, so it's doing something here where it's just like previewing. It's It's iterating upon. So I believe this would probably be the start of the skydiving logic. It's funny. The thing is like we just wouldn't normally see this, so we can't laugh at it fully. It's just the unexpectedness of it and this clawed orange cube, I suppose. All right, so here's our skydiving result. It was running it in a preview server there, but it wasn't loading, so I had to yell at it to stop it. All right, drop zone. Maybe drop zone or drop zone. It doesn't matter. Exit the aircraft, track across the sky, pull at the right moment, then land on this soft target. Okay, shift is to flare. All right, good. It drops us out of a plane, which the Opus 4.8 and the GLM 5.2 did not do, which was oddly frustrating. Okay, the problem is the camera is not really properly following us, and it seems like we actually have a decent player model. It's just um Oh, okay. Maybe we do. May Okay, maybe shift to flare. This is going to be so slow. All right, I just went out for lunch and now I'm back and we're still uh in midair. Now, these games are always like this. I'm going to say the clouds are rather dense from what we can see here. The terrain actually seems quite wellformed, especially because the comparison of the Opus 4.8 result is fresh in my mind. It's just uh it's very slow. That was my own fault for pulling the parachute too early. Let's try again. Okay, the plane looks good. And the wind noise actually is variable depending on how fast we're going and if the parachute is deployed or not. I'm going to just not deploy the parachute or wait till like we're right near the ground and then we'll see. Big issues are the camera doesn't actually show our player when we're just diving. Okay, we kind of got a view there and it looked cool. I would say just from like right here, this looks okay. The terrain is well done. It's a good mix of water and like nonwater. So land and then peaks and things like that. What happens if we just don't deploy the parachute? So oh it automatically deploys for us. What an anthropic thing to do. Okay. The other results even oas 4.8 and GLM52 did not um didn't autodeploy. Oh, all right. Hard landing. And it gives us a bunch of statistics and things like that. Auto deployed, low, penalty applied. Next up, I'm going to be giving this a front-end design test in the form of the beautiful website for Slapis Watch Company. It should feature a high-end hero section with an animation panning around the watch, a 3D model of a watch, and it needs to have like a cinematic pan in the hero section. Additionally to that, it needs to create the 3D models of the watches by itself. So, we get to see a little bit of its 3D modeling and design chops as well as front end. Usually, these go for like a black and gold elegant theme or something of the sort, but I'm more interested in seeing how the camera animations and the 3D models of the watches look. All right, I'm blocking my vision right now cuz I don't want to spoil it, but here is our beautiful watch website for Slapus Watch Company. Okay, this has some potential. We have our azimuth location. Make of that what you will. The watch is wellformed and it looks okay. the face. The hands are oriented correctly in the center of the watch. It has chosen a time of well I suppose that would be 720 but it's too close to 8 to be 720 but it's too under eight to be 820. So you know it's something quiet by design. We do have our crown and we do have the straps. They are very very narrow comparatively to the watch body. We'd rather make two watches than 20. So, an emphasis on a low volume perhaps boutique watch producer. All right. And then we have the watches here. One with a blue face and one with a gold face. The Riviera and the Meridian. And the strap for the Meridian, excuse me, the strap for the Riviera is like a solid leather band versus the aluminum band for the Meridian strap is too small. I think if it had properly sized the straps here, this would be a significantly more impressive result visually. But this is something I always enjoy testing because it has to do front-end design, 3D modeling, and then the ability to understand what a cinematic keyshot render in the hero section would look like. So, overall, acceptable. Next up, we're going to be running this on X high. That is one below max. And I want to get a feel for how this does in a more balanced time to performance setting, which would be XHigh. This is going to be a new prompt as well that was first introduced in my previous GLM versus Opus video. Create a 3D scene of a city block. Emphasis on detail is very important. The scene must have a timeline slider in the top with the following options. And then we have the options of five specific years 25 or 20 years apart each. So 1945, 1965, 1985, 2005, and 2025. The point of the scene is to be able to select any of the five different years. The scene will transform right in front of your eyes to the time period selected from the slider. The time period should affect all aspects of the city block, the buildings, vehicles, storefronts, advertisements, outfits on pedestrians, everything. This must be a polished high-end scene with sound effects, ability to navigate around and look at things, etc. Go all out. So, we'll see how this does just on XHigh instead of Max or Ultra Code. All right. So, we have our timetraveling city result here. Okay. Timeline city block. A city block through time. Enter the street. That oddity to the camera is quite a letdown because this result actually looks like it may have some potential. Interesting. Opus 4.8 also included the Beiju Theater. We have some pedestrians walking and take note of the cars, the building styles. First National Bank, the street lights, things like this. I I always keep trying to click when I'm on these. We have a street car on a railway, although very odd. Now, let's see if we can go to 1965. Good. The way it kind of just goes up on the buildings is interesting. It's a little more happy. However, okay, Metro Savings Bank, we have some lit up buildings on the outskirts. Now, the cars are more flamboyant in terms of their color schemes. palace cinema. The street lights. We have some trees. Let's see if we can find pedestrians to take a look at the Will they Oh, okay. All right. 1985. Oh, wow. Yep. This is exactly what one would expect. Super synth wave and police car siren sound or something. Interesting. ElectroMart video rental chrome diner. Press V to switch to a drone camera and see the whole block. Ah, okay. Good. I like that it showed me that tool tip. Then 2005.com hangover. Yeah, that could be said. Multiplex. All right, let's check out our pedestrians. Hey, get back here. All right, fire hydrant. Cyber cafe. The street lights change. Everything's a little more toned down. All right. And then finally, 20th, 2025, smart block. Very similar except for the emissive properties of the buildings, making them somewhat blinding. We still have a cinema quickart and then the cars. It didn't change the headlights on the cars, which were seen with GLM52 and Opus 4.8. It turned them into like LED strips. Okay. Overall, this is a troubled result. It has I mean it's a hard test I think but the visual artifacting that we're seeing here is a little upsetting. The drone camera capability to see that is pretty cool as well. So if we just restart it from here we can see there's a lot less visible detail just because the buildings do get taller though. Notice that. That is something that is a nice attention to detail. And we can zoom in around and pan around. So, and they do have smartphones in their hands in 2025, which is something interesting that gets included. So, I'll do it one more time here. Just cycling through these from this angle. I just love the 85. I love synth wave. I have no idea why. I wasn't even alive then, but I like the color palette. Is that person running? No, it's uh and they're on their smartphone. So, interesting. Troubled result for 40 something minutes, but nonetheless, a new test to include. So next up, I have just put this on extra thinking which is one step down from max and I'm giving this a test that is kind of a multimodal coding test. Oh, that was silly. I did put it in the folder which had the photos. So let me the photos are in this directory. So the photos are just of this 3D printed laptop that I've created. So, there's a bunch of different angles here, and we wanted to replicate that in 3D along with including a keyboard that can actually be typed on just by clicking on the keys. All right, I've stopped this after like 40 something minutes and said, "Just give me what you have. I can't take this anymore." Yes, it's unfair. The result won't be as good as it may have potentially been able to be at this point. I'm fine with that. I just want to see what it would have done. Inevitably, this will take a bit longer. This is on extra mode, so it could be thought of as my own fault. But, you know, it's just like so incredibly slow beyond what I would have expected. All right, so let's see what it would have made for the You know what? That's actually not half bad. This is got some potential. Yes, it was troubleshooting the issues here with the faces not appearing properly, but I will note that it did have the switch and the port in the proper positions. Something that still to this day I don't think I've seen is the lid is supposed to be like the hinge is supposed to be over here. None of them really get that. It did write 1 ghahz turbo clock. The keyboard is backwards and inverted which happens a lot. Okay, we do have key presses. I just need to do it backwards. So help. And I do like the color like additionally on the keys. This is actually not as potentially scary like bad as I thought it would be. Okay. Unfortunately, it doesn't seem like that works, but it did do a good job replicating the screen and the shell that would be seen in it. Open close lid. Okay, not bad. Well, within the confines of that. And then power to turn it on and off. Okay, given infinite amounts of time and money, this may have produced something respectable, but I needed to stop it at that point cuz it was like, all right, so I'll give it something random here just on medium to make us a 3D low poly Formula 1 racing game. It should have a camera chase view and then a driver camera view. So from in the car versus a chase cam. Make sure it looks good and fantastic looking and fun. And this is on medium just to get a feel because I am interested in seeing how this performs across a few different potential reasoning effort levels. All right, here is our medium thinking level simple formula 1 racing game result. Apex circuit. Oh. All right. I will be honest. I expected this to be a little better. The steering is inverted. Off track. Oh well. my mistake. All right. And let's see. How do we change our camera view? See? Oh, okay. Good. So, I did specifically tell it that it should look very good from the driver view. This looks like a I'm about to be born view. Okay. And we're off track. I'm sorry. This has just not really been what I expected. Truth be told. Okay. And that's that's a it's a it's a camera toggle view. Good thing is the wheels are oriented correctly. They do spin. The steering is pretty visible and things like this. It's just and this was on medium thinking. So just keep that in mind as well. I've put it on high here just in the web chat interface because I have things running in parallel across cloud code and the new desktop beta app. This is going to be the drum kit simulation test with the autoplay feature just to see how it does in terms of orienting the drums if nothing more. Is that's something that is still difficult to all but the most capable of models. Okay, this is solid. I'm definitely okay with this on first glance. Let's turn the speaker up and jam some beats since there's nothing else to do but wait. The high hat movement is interesting. The symbol is unfortunately not visible. At least the one that would be on the top there. All right, let's try the autoplay feature. A bit awkward, but functional funk groove. Hip-hop. That one slaps. Then jazz swing. Okay, acceptable. So, I'm giving this a test that is including an Opus 4.8 result. In specific, the Opus 4.8 skydiving test had some fantastic terrain. So, I've given this access to that folder using Sonnet 5 now just on high effort and have told it to keep the terrain from that result and turn it into a simple driving sim with an off-road vehicle soft body suspension. So, essentially articulating suspension that will travel over that terrain. It'll be interesting to see how this does when using the basis of something that Opus 4.8 made and then having to strip it and turn it into a new style of game. Okay, this was the final thing I had done where we had given it the existing result that Opus 4.8 made for the skydiving game and said turn this into a low poly off-road driving game with an off-roader and like some soft body physics. It didn't fully finish, but I said just give me what you got. Okay. The sounds are not bad. Okay. Uh-oh. All right. Well, it was just something I wanted to do just to see how it did compared like when working off of the sick terrain result that Opus had made. So, I am going to work on that on my own time. But that's going to bring us into the conclusion which I've been eagerly awaiting because this has taken a very very very long time. Overall, I have to say honestly I'm not impressed. I don't know what I was expecting. It is a Sonnet class model which is in the middle tier of intelligence of the publicly available anthropic models at least currently. It feels like a pretty small model just in terms of size. I have no justification for that beyond the vibes I felt while using it. But I noticed it uses it's just so incredibly slow to use on any decently capable thinking level, which was kind of a let down. It had some capabilities. I mean, it did all right with some of these tests. I I think the big at least in terms of like what folks are going to see on X or talk about the big question here is going to be genuinely like does this out compete not GLM52 cuz that's not necessarily a direct competitor pricing wise it definitely could be but things like cursors composer Gemini flash 3.5 flash I don't know that this is going to really be a decent option for most use cases when there are so many other cheaper and more performant models. This seems to it just uses so many tokens. I mean, this was cool. The auto like that model right there looks pretty good and the auto deploying was kind of funny. So, I I find that genuinely I've come away from this test honestly a bit frustrated because I I I don't know. I just I expected better to be completely honest. It is competent, but that's for the pricing that it's going to be. I don't really look into like like initial pricing half off. I don't care like because eventually it's going to be full price and looking at the post August 31st price for this. I don't know how I feel. So, I'll be interested to see what folks think of this model. But that's going to conclude our first look and test of Claude Sonnet 5. Thanks for watching.