GLM-5.2 vs Claude Opus 4.8 – Does GLM REALLY Beat Claude?

So, an interesting sideby-side comparison of some of the capabilities. Today, we're going to be putting GLM 5.2 in a head-to-head comparison test against Claude Opus 4.8. So, as we can see on the screen right here, both of these models are currently running in their respective harnesses performing the same task that I've given to either of them, and we'll touch upon some of these specific testing methodologies in a little bit. But the main point here is that I've heard a lot of popular discussion lately talking about how GLM52 may be the first open weights model to either match or exceed the capabilities of a state-of-the-art closed model like Opus 4.8. However, the thing that I noticed that I didn't like is I see a lot of claims being made such as that, but none of them really have any photographic or visual evidence to them. So, I wanted to just give these the same exact test so we can objectively compare the results side by side. Now, in mentioning that, there are a couple of additional things I would like to touch upon. One being, please feel free to subscribe so I can get that 100K plaque. But two, I want to showcase prompts that have never been done on this channel before across a wide variety of different programming languages and tasks. So in lie of kind of some of the more fun and pretty 3JS web results, I want to test these things across some other languages that may be a little more boring visually, but will hopefully showcase the capabilities or lack thereof of either of these models across a bunch of different and perhaps less commonly explored domains. So in terms of how these are being tested, I do want to make note of that because I know everybody has their opinions on what specific coding harness is best for what model. Because of that, I've opted to use either of these in their respective harnesses that come from their own manufacturer. So, ZAI right here is being run in Zcode. This is using GLM 5.2 with a million context window on max reasoning, which is its highest reasoning level. I've opted to use this because, well, it is somewhat similar to Codeex aesthetically, but they do offer this and say it's specifically optimized for GLM52. So I wanted to use it in something that was and for claude it's just makes sense to use it through claude code. Unfortunately there is no officially supported UI desktop or guey application for Linux from Enthropic. So we are stuck with just using it through the command line. Now in terms of the specific subscriptions which are being used for either of these I am using the middle tier GLM coding plan and it's around I believe $65 a month. I don't know that I'd specifically recommend it because we're probably going to notice a lot of reconnecting things that come up here during the process of running these tests, but regardless, it gives us access to this and a reasonable usage allotment, I do believe. For the clawed usage, I am on the $100 a month plan right now currently, which would be the max 5X. So, both of these are just being used through their respective subscription plans. All right, so now that we have our results, we can actually look at the initial prompt that was sent, which is very simple and open-ended on purpose, at least for this first sidebyside test. The prompt was to create a 3D skydiving simulator. This is deliberately an open-ended task. The exact implementation of the simulation 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. Now because both of these has finished and interestingly they both took very similar amounts of time. GLM 5.2 took 26 minutes and 57 seconds and Opus 4.8 took 27 minutes and 39 seconds which is interesting. There's no correlation. It's just kind of funny that they took right around the same amount of time. So let's start by taking a look at the GLM 5.2 result. If I click on this URL right here, I'm just following the instructions that were listed right here when it gave us the final generated result. We should be able to just control-click this and then see our skydiving simulation. Okay, edit pose. It is possible that is a skydiving term that I'm just not familiar with. So, let's take a look. Okay, jump. Interesting. We definitely have the sensation of speed. I did also tell them to put sound effects in. Let me turn my speaker on. I think it's possible that I may have misconfigured the audio settings, which I will fix. Oh, so we have to land there. Now, my thing is one would normally try to get a parachute. Okay. Yeah, I wasn't How do we play this controls? All right, let's try that again. I do have the sound working and I also have figured out. Okay, that's a good sound. So, we can also toggle camera. Okay, and we have our individual. Here we are. We also have W. I believe will speed up and then S is to slow down. Now I don't know when it would Oh, deploy shoot. Okay, nice, nice, nice. That's not bad. You know that all of the strings from this chute are actually connected or at least showing towards the body of our diver here. And it seems like the mouse can also affect the lean and movement that we have. And now we're just kind of like gliding. There was W and S. I think W would allow us to speed up. Okay, S might kind of is pulling up on the shoot. So, you go up and then W is to go down more. Seems like we're going to glide down pretty nicely. And I also have noticed that the aggression aggression of the sound of the wind noise is a bit calmer than it was because we've slowed down with the chute deployed. Let's change our camera real quick. Okay, cool. All right, we're about to Yes. Yes. Yes. Fantastic. All right. Clean, greased it. Rings. Oh, we're Yeah. All right. So, I'm not going to go through the entire thing, but do you see these floating rings here? So, apparently we're supposed to fly through those, which adds a level of complexity here. Nice. Okay, cool. All right. So, that was what we were supposed to do. Do you think we can just like let go of the shoot if we press space? Okay. No, we can't. Good. Good. We wouldn't want that. That was well done. And that was the GLM52 result. Good. And it act it implemented some gameplay elements such as getting through the rings, which is what we wanted to see. All right. Next up, now let's take a look at our claude result. If we click right here, we will get it. Okay. Preparing the jump. Skyfall. Exit the aircraft. I will just take a look at the controls. Space is to deploy canopy, which would be parachute. Okay. A and is to steer. S is flare to land. All right. Let's check it out. Interesting. So, they both have very similar sound effects. This is definitely actually that terrain map is is incredibly well done. And we have a heading thing on the top like a compass. with what degree. This is pretty good. They both have very similar implementations for the wind noise, which is to be expected, I would think. I don't know what this like diamond thing is right here. I wonder if that is a Yes. Okay. So, that's pointing us towards where we're supposed to land. We are still in freef fall. We have an altimeter, but the actual different color patches of ground here as well as the terrain. I want this map turned into like a a low poly like off-roading game, which I may do on my own time, but okay, we have our person. There is a shoot backpack. Okay, deploy canopy. I mean, I I want to see what happens if we don't. We'll try that first and then in the next one, we'll uh we'll we'll do it. Pull deploy now. Do you think it's going to Nope. Okay. Oh, all right. Good. Now, we'll we'll try that again, but this time we will ensure we do that just for completeness of the test. Very nice map, though. There's even a little lake over there, so it's not just the same patches of terrain replicated. Okay, nice. And this also does have rings that are connected to the player. We also can like pull up to slow down as well. And that is actually just an effect on like a it makes the parachute a bit bigger, I suppose could be said. It scales that model to larger. This like crosshair right here, I keep instinctively trying to like wanting to click the mouse because it just reminds me of one of our FPS tests. But that's okay. Let's take a look at this landscape real quick. Very well done. Very pretty landscape. Neither of them included the plane, though. They just started us in midair, which I think is a pretty like abhorrent emission of a gameplay asset that one would expect, but you know, what do I know? I don't think we're going to hit the mark. I am holding it. All right, we're not going to hit our mark, but Oh, unless it's assisting us now because we were not making much progress there. Okay, I'm holding S to flare. All right, hard landing one star. That was very, very lengthy of a process, but overall this was that terrain down there is if I had ever received that in one of our flight simulator prompts, that would have been an automatic insane in the title. So, an interesting sideby-side comparison of some of the capabilities and something that I wanted to do that I've never ever done on the channel. This came to me randomly cuz I have no idea what causes human cognition and thought. So, with that, let's now go ahead and do something a little more difficult. So, for this next test, I am going to put Claude on the max effort setting. We were on X high previously, but I want to do full-on like head-to-head intelligence comparison for this one. I'm not going to use ultra code and I'm going to just pretend that one's not there for the purpose of this. But our next test is going to be a bit more involved because the ultimate goal here as we will see if we expand this prompt is to create a working Windows XP compatible guey chat app and provide a built-in .exe file. This is going to essentially be an AI chat front end which is connecting to a system behind me that is serving the Ornith 35b model. Basically, we want a little LM Studio for Windows XP. To put it very simply, we have given it specific information like the shape of the request that it needs to send the specific endpoint. It is saying the laptop will be hooked into the LAN. So, hardcode these values in and then give us a good-looking chat interface in an .exe file basically that works on this laptop that will allow us to actually have an AI chat app on the XP laptop. All right, so we have both of our results here. GLM 5.2 finished in a little under 23 minutes and Claude Opus 4.8 on the XH high thinking mode finished a little under 28 minutes. Both of them interestingly pointed out kind of the same restrictions in terms of their level of troubleshooting was stopped because this Linux box does not have wine. So they couldn't launch the actual GUI to eyeball it. And interestingly, just good to see that they both picked up the same caveats of being able to fully troubleshoot this or QA check it from end to end. All right, so unfortunately my normal testing XP laptop does not seem to be working right now. It won't charge or even accept any power. Nonetheless, I have a little XP netbook that I've just transferred the files over to. I'm going to record the screen of this with my other phone, which has a pretty abysmal camera, so I do apologize for that up front. But I've transferred over both folders containing the exe chat interfaces. Now I just need to connect this to the LAN that the box serving the model is on and we'll see if these work. All right, let's start with GLM and hypothetically this should open. Okay, so it did open. That's a good first step. Let's do new chat. Okay, let's just send a message and see if the system responds. Hello Okay, we have to manually ah okay so unfortunately the GLM one does open. It's pretty simple but we are getting an error immediately. So let's try the clawed one and see if there's any difference in performance. If not it is possible it's something systembased that is blocking both of these. All right. So here's our clawed result. Very very similar in terms of the aesthetics. No conversations. Type a message and press enter. Hello. could not reach the server. The system cannot find the files specified. So, interestingly, both of these worked in that they did properly make a simple guey and both of them launched via the .exe on the Windows XP laptop. Unfortunately, neither of them were able to successfully either send or receive a response from the local AI server. In the case of GLM 5.2, when we tried sending a message, it basically said Windows has encountered an error and then it needed to close. With the Claude app, we sent a message and it just said, "It can't make a connection. Make sure the network settings are all correct." So, neither of these really worked. So, they've both been given a follow-up prompt here just to massively simplify some of the ways that the connection works and things like that in the hopes that we do get something functional as a response from our local AI server. Both models have finished the follow-up to try to get the chat app working for Windows XP. Interestingly, Claude was a bit quicker this time than GLM. GLM was 15 minutes and 50 seconds. Claude was only 9 minutes and 31 seconds. All right. First, we'll try the GLM chat. This is the fixed one. And we'll just do hello and then press send. Darn it. Okay. So, this one's encountering an error. Unfortunately, I was hoping maybe it would give us an error instead of quit, but all right. Nonetheless, let's try the claw chat one now. Interesting. And it automatically tried a connection check. Let me do a new conversation because that could be left over. We'll do hello. Interesting. All right. At this point, I am regretting this test, but both of them have at least given us some bit of error messages that we can send because neither of them worked this time around as well. So, I have the specific error that Claude had given us from within the chat interface. So, I'm going to paste it in here and just say fix this. And we don't necessarily have an error message for the GLM one. It just has the Windows encountered a problem and needed to close. I am just going to give it what was listed from within that error message. It may not be much to go on, but we'll give it the same chance with the bit of information that we do have and we'll see if we get a working result from either of these. I think at this point this will be the last chance that I give them. All right. So, these have both hypothetically fixed the issues they were facing. Now, I want to be transparent and say I may have made a mistake here because this is the third try for both of these. However, I incorrectly assumed that the reason the GLM one was failing was because it wasn't making network access. I now know that was mistaken. So, I likely should have pasted in this error code that was appearing in the second test we ran. So, this is the third try. I should have pasted that in the second try. So, if the GLM1 has a network issue as Claude had found it had here, I may give that another chance to fix it because there is some level of human error there in my incorrect assessment of what was causing the issues. All right, let's start with the GLM chat now. So, we should be able to send a message without the crash that happens previously when we had sent it. Very good. And now it does not fail when we send a message. Now, it is questionable whether or not we'll actually get an assistant response here or some form of error message. Fantastic. That is actually quite impressive because this likely then Well, okay, let me try clawed and then we'll see what happens. So, this was the GLM chat result. Now, let's try the clawed one. All right. So, now we're trying the clawed one. And it seems like both of these are going to work. As I look over at the video card there, it is being utilized now. So, we do have two functional chat applications working on Windows XP connected to a remote local AI server. Well, on the LAN, so very cool. All right. So, I'm going to be honest about this one. I am actually a bit surprised because these both got three attempts to do it. The thing is, it seems like the issue that the GLM one had was that it was trying to include something. So, I guess what I'm trying to say is the GLM result seemingly was including something related to rendering text that was just causing it to crash when we tried sending something. It didn't actually seem to have any network issues unless it had resolved it after the second chance when we told it incorrectly that it was crashing because the network was not properly configured. So, the thing is the clawed one needed more work to actually get the network aspect of this functional. It seems like the GLM one just needed to be fixed because it included something that wasn't working in terms of like text rendering, which is kind of interesting. So, this is definitely going to be an ask the audience one, but just an objective sidebyside comparison test where they both got three tries and they both ended up getting it working. though the tries culminated in needing to fix different issues to get either result working, but the results did work. So, this was a very, very interesting one. So, next up, we're going to be giving these a different type of test where they need to create a 3D model that we're actually going to be printing with the bounds that it must be fitting of a real physical object. So, they need to know the dimensions of that physical object as well. The prompt says, "I have an N20 motor, which is a very small little DC motor used in tiny RC cars or maybe tiny robots. I need you to create an STL, which is just a file that can be 3D printed that I can 3D print that is a V8 engine and transmission combo that the little motor will fit inside, giving my RC car project a realistic looking engine that houses the actual motor. So the process here is to one understand the sizing and dimensions of this motor and then two create a decent looking model of an engine and transmission that the motor actually fits inside of. Considerations for this will be that the motor has an output shaft. So there will need to be a hole for the output shaft to go out. It also has two connections for the positive and negative connections for the motor. So there will also need to be some thought put into allowing access for those wires for the motor. So we'll see what we get here. This will be interesting and I would like to actually 3D print the results so we can physically see if the little motor fits in them or not. Interesting. So GLM actually asked us some follow-up questions where it's saying what is the specific type of N20 motor. Sometimes they have a little gearbox attached to them. In this case, this motor does in fact have said gearbox. So it is good it asked about that. So I will just select the standard geared N20 right there. I would assume that Claude will probably just assume that it's the standard geared version like this, but it'll be interesting to see. And again, we have another question from GLM just asking specifically what this composition should perform. So, is this just going to be a static model or does the motor actually need to be able to drive wheels? So, like should it put a hole in for the output shaft, etc. All right, it's asking another question just at what direction is the shaft going to output? So, it could be for a rear wheel drive car, for a front-wheel drive car, or for something else. We just selected the normal longitudinal style, which would be traditionally seen in a rear wheel drive car. Okay, it's asking some additional questions like what style did we want? I just said classic LS style V8. What material are you printing with? Good. So, FDM, which is just regular plastic, like your off-the-shelf 3D printer, not a stinky disgusting resin printer. All right. So GLM is still asking me some questions like how do you want me to do this handwritten binary STL? Yes. No, I'm not going to I'm not going to have it do that. It I because Opus is just doing however it wanted to and it does seem to mention OpenCAD. I'm going to say just do what you want. I just want the STL file. I am a little concerned because GLM the counter for how long it's been working is still iterating but it's like I don't know what's going on right here. It said thought for a few seconds and that was more than a few seconds. It was thinking intricately about just planning this all out. I at some point was just looking at it as it was going through its chain of thought here and it was like gh but how does a V8 orient? And it was like oh okay so writing now you've given me everything I need and the latitude to just build it the user instruct into just give me the STL file overrides the specw writing ceremony of the brainstorming skill indeed it does. I'm going straight to building with a render verify loop to make sure it's actually correct. It should have given us some feedback that it was actually doing something. I think that it's possible it had frozen up or something. Their API access is not the most reliable. So the claude result has finished here. The GLM1 is still working. This does not natively have vision, the GLM52 model, but it is seemingly working around that in some way, which is fine. It's using a tool to be able to assess what's in images, and it's just saying, okay, the problem is that the V is inverted in this image. In the meantime, we do properly have our claude result. So, I want to just take a look at that in a 3D printer slicer, and then I'll probably just print them one at a time. All right. So, I'm going to open this in my printing slicer program just so we'll be able to get a look at it right here. This is the one that has the full plate. Okay, this is let's see. Let's I have to remember how to and then we have a slot for the motor. So, this is this is rather frustrating because while very very very well done from an aesthetically pleasing standpoint, the problem is we can see right here there's a slot for the motor. This is upside down here. Let's take a look and see if we can see any of these renders. The problem is there's no output for the motor shaft. There's no hole there. and it did not account for the fact that the N20 may have had a gearbox on it, which many many of these do. And it didn't ask us. That's very frustrating, I'm going to say. So, here we can look at the result that Opus made just in open SCAD. This looks pretty good. It is definitely a V8. We have the air cleaner. We have the valve covers. We have the transmission. The big problem is, as we saw before with the split view and seeing where the motor was supposed to go inside, there is no actual hole for the output shaft. So, the motor would not actually be able to be used. And I did say it is in a remote control car use case, which would be indicative of the fact that the motor would need to be used. So, it did put holes for the wires to go in. And additionally, the issue was that the gearbox was not given. Oh, unless it's possible that it made it long enough that the gearbox on the motor would slide in there. But the issue still is that the motor shaft has no outlet. Now, in the meantime, GLM has just swapped to completely speaking Chinese. I'm okay with that because the end result is an STL file. All right, so we now have our GLM 5.2 result as well. So let's just start by taking a look at the 3D model. Interesting. So this created a bunch of different models instead of just one that we could print. So let me open this in OpenCAD. It's not it's difficult for me to assess this off the bat. So I need a moment. So let's just look at some of these parts individually in the slicer program. So, okay, this was titled block. This is seemingly pretty simple. It doesn't seem like the motor is actually going to be placed in here at all, which is fine. It is can be an aesthetic piece. Next up, we have our headers for left and right. Yes, and those are mirrored. Good. So, left and right does work. Intake. Ah, that's actually not half bad. I would not want to 3D print that, but that in and of itself is not bad. Transmission left. Now, this should be where our motor is going to be placed within. And okay, maybe not. Maybe it doesn't go in there. Let's just look at our right transmission, which is that just mirrored. Then we have our valve cover. Okay, so that's what the funky spiky thing is there in the open scad thing. Okay, so really I guess I would say the main issue I see here is that the transmission seems to be where the motor is supposed to go in. So mainly this piece and then the other one right there which is just the other side of it. The big issue is these floating cylinders right here would completely inhibit the motor from fitting in there. Additionally to that, it also still neglects to have the output shaft. So the output shaft, there's a wall right there which would stop it from going out. And then these pieces don't actually seem to be attached to anything. So, they would just try to print in midair and that would likely be okay and you could just pull them off after, but it would still cause a bit of issue. I do see just in the back right here, there is a little space for the wires to come through. I would assume that's what that was. So, it's kind of interesting that both of these had similar failure points in that the actual output shaft of the motor has no outlet. They both did put in consideration for routing the wires. Unfortunately, just because of some of the issues we see with this one, this really is not worth trying to print cuz we can see there's some pretty blatant issues here. And the motor wouldn't necessarily fit properly in there. But it's just interesting nonetheless. And this is something in specific as a test where Fable 5 really really was a different species of model almost because its capability was so significantly improved over any existing model in creating STLs that can be printed such as this one right here which was designed to fit a much larger sized RC motor. And this was just a totally totally different league than any other model was the engine model that Opus 4.8 had made designed to fit this tiny little electric motor. Overall, it looks pretty good. But the big question is, does it actually fit inside? Did it accommodate for the gearbox? So, let me just try to Yep. Plenty of room there even for wires to come in. So, aside from the fact that the output shaft does not have anywhere to go, if we sandwich these together like this, that is pretty cool. But still, this is somewhere where Fable definitely was like a different species of model in its ability to do stuff like this. But this is quite respectable. Our final test is going to be somewhat of a 3D asset, but not necessarily a game. So, the prompt is to create a 3D scene of a city block. Emphasis on detail is very important. The scene must have a timeline slider at the top with the following five options: 1945, 1965, 1985, 2005, and 2025. The point of the scene is to be able to select any of the five different years, and 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, it'll just be interesting as not only do they have to create a competent 3D scene, but also I want to see how they handle transposition of time periods between selected years. if it actually shows any cool visual effects as things are changing from one year to another or if it's just more of a static flick of the switch style thing. So, it'll be interesting to see how they tackle this and how they stack up. All right. Unfortunately, GLM like paused again. This is what happened with the 3D model where it seems like when it's asking follow-up questions, it just it's very buggy in this app. I think this app is still in beta for Ubuntu. However, at least we do have a gooey desktop app. All right, it didn't create anything in there. So, I'm gonna have to make a slight change to the prompt here, telling it, "Don't ask me any follow-up questions. Just build it." All right, after about an hour, we finally have both completed time travel or time scene results. I'm going to turn the speaker on here and let's just start with GLM. It does say it is opened or running in the browser. So, okay. Krono City, one block, five eras. I thought that said five ears. I was like, "Oh, okay." Travel 80 years in the drag of real. Watch a single city block reborn across the post-war smog of the 40s, the neon swagger of the 60s, the chrome glare of the 80s, the glass towers of the 2000s, and the quiet glow of today. All right, let's enter the block. That's pretty cool. Okay, so this is Oh, look at the walking animations. At least from up here, that looks really good. Drag to look. Scroll to zoom. right to pan. Okay, the cars look cool. And this is the 1940s, so we can just give me a second here to like orient myself just as like a city scene. This is not bad. The cars do look rather vintage. And the clothing on the pedestrians was something we had specifically talked about needing to be period correct, as well as the storefronts. So, this individual has a top hat. And they do seem to, at least from what we can see, be dressed in more fancy clothing. All right. Okay, that's that's going to be a traffic violation there, but that's all right. Mood somber, smoke and soot. Okay. What if we do auto tour? P is to press. That was sick. Okay, so auto tour changes the time for you. Okay, so now we have 85, the neon decade. We're going to look a little more intricately at these, but the actual changing between different eras is very that is like total like stale 2005. That's okay. So that is what the auto tour was. That's very very cool. All right, the animations between them are cool. Look at the fins on the cars. Absolutely nailed it for 65. The colors. Look at the change in like even the lights. Do the street lights actually change to reflect? Yes, they do. Very interesting. So, this is more like mid-century modern style. We have some cool aesthetic to like all the colors, the individual walking animations, the flared skirts, the fun colors. This is pretty cool. And I love the way that it changes the scenes. It doesn't just statically like I'm impressed with this. All right. 1985. Okay. Everything is very very synth wave. The cars are more defined in like blocky shapes. Let's see. It's hard to get a good visual on the outfits because it did change the color palette so aggressively, but we see. Okay. Music TV right there. Play games. the the F club. Uh, okay. We maybe just ignore that. And then we have uh I wonder did we have visible shop signs here in the 60s or the 40s? Okay. Bit difficult. Hey, the street lights actually changed. So, that's cool. All right. Let's check 2005. Cool. The cars are more bland and less exciting, which I think makes sense. Everybody's in monotone colors or something. I don't know what the clicking noises, but that's all right. Do we have any visible signs? Okay, dental, go digital. I would imagine that would be sushi. And then people have backpacks on, so that's pretty cool. And the street lights have also changed again as we can notice right here. All right. Then finally, 2025. Cool. Everything's a bit more colorful. Definitely more similar visually to 2005. The cars have like LED strips. Look at the way that the headlights changed on the cars for 2025 as opposed to everything else. As well as the street lights, more LED. We do have some interesting sound effects here. And oh, I can move around that way, too. I apologize for that. But that was a better place to look at everything. Cloud first. Okay. Vape. Yeah. All right. Got that one right. And then what do we have over here? Ride electric. Solar panels. It even changed like what would be on the building top there. That's cool. And it actually has smog particle. This was very, very, very well done. And this was the GLM52 result. Incredibly well done. All right. And now we have our Opus 4.8 result. This was run on X high thinking. So we'll see how we get it. Okay. Chronoblock. Did they Okay, that's that is very very good. I was going to say, did they name it the same thing? What? This is post-war optimism. Late afternoon VJ day still fresh. Brick walkups, painted wall ads, and the clang of a street car. Men in fedoras, women in tea dresses, chrome bumper sedans gliding past corner markets. Yep. The outfits. And can we move up and down? Yeah, we can. Okay. I mean, this is Wow. This is very good. Shave 5 cents cigars shaving the Look at these building textures. We have some like hum the street lights are more like old-fashioned. All right, I want to click tour because in the same way that we were able to almost accidentally see how this was. I'm just going to Okay, that's so cool. They did such a Look, these are very, very similar. 1985 is depressing. 2005 glass towers, more modern, and then 2025, perhaps a bit optimistic in terms of the, you know, they had very, very similar changes between the years. So, okay, now we have the I don't love this like ocean sound that it's put in, but I'll leave it going just because So, the cars have fins. All of the colors have become far more vibrant as we noticed before. The street lights also did change between the two of them. I find that interesting just seeing the attention to detail in that as well as the storefronts. Does that say BJ? No, it doesn't. It says Beiju, but one can dream. All right, I'm going to turn the sound off because it's kind of a pain. That's the Beiju cinema. The cars have fins as we saw with the other result, too. Cola ice cold refresh. And that's what we get there. The buildings also have changed to be a little more optimistic and happy looking. Let me zoom out. And the bus changed as well, more in line with just its color. We have our bus stop there. Now turn it into FPS. That's I don't know why that's like one of the things that comes to mind. So, all right. Now, let's take a look and I will turn sound back on. Let's go to 1985, which was interesting because this is like kind of depressing. I don't know why 1985 just became really sad. The folks do have colors on that one would expect to see here. It's pretty much the same sound. The street lights I think have changed again a bit. What are our ads? I do believe they said like uh Walkman personal stereo. Makes sense. The buildings are like a bit more corporate and less flare I suppose could be said. And the cars are just interesting horn sound. All right, let's take a look now at 2005. Much more optimistic. We have the same sound effects. The cars between 85 and 2005 are not necessarily much different visibly. There's like more truck looking there. There's a lot of corporate business folks walking around here. Something I noticed in the GLM one was they had backpacks, which was kind of cool. Okay, we have the looping horn. The bus says downtown, or at least it's supposed to. This is just this is cool. All right. And then we have 2025. Seems happy and optimistic. The street lights. Interesting. Do you notice that between 2005 and 2000? Oh, no. It was between 1985 and 2005, the exposed power lines actually went away. So, that's interesting to see. And if we go to 2025, we have some more like eco-friendly and green things. What is that going with the person? I almost wonder if that's supposed to be like a smartphone. Maybe it's that they're all being tracked. Those are those are tracking. No, I'm kidding. That's messed up. Okay, this person is I mean, the clothing is more like calm. A bike rack with bicycles as well. And interesting, the car headlights have also changed with this one as well to be LED. The behaviors here are very, very interesting to see what things are consistent amongst different families of models as to what they keep. Even to the point where we have some rooftop foliage up here as well as solar panels. Both of these really did include a lot of similarities in a pretty like open and free prompt here. So, this was just I mean I wanted to do prompts and tests that were never done before on the channel. This scene right here, I think this was probably the most impressive scene for this specific test was the 1945 from Opus. The buildings were very personal good personality. I've been at this for quite a while. Both had such similar like translation animations there in terms of time. I mean, that's really that is really quite cool. That is going to conclude what took a bit longer than I expected, but a proper sidebyside test between GLM 5.2 and Claude Opus 4.8 run on either XH High or Max. I'm going to say the differences in capability between the two are quite stark when it comes to like 3D asset design like 3JS things. Claude definitely still has a massive lead. We saw a bit of a lead. I mean the GLM STL result to basically make the motor thing was respectable but not perfect. Claw definitely outperformed it there. Something I'm very surprised at was the Windows XP AI chat app test. I do believe GLM52 may have actually had that working sooner if I had not messed up in my troubleshooting. So, they both did it in three tries and it both worked. The thing is GLM 5.2 had a simpler issue that was blocking it from working. That would have been resolved in the second try for that had I just given it the error log instead of assuming that it was crashing because no network connection was being made. Claude had a bit more of an intricate issue, but they both did ultimately get it working and it was such a simple interface that both of them were pretty much identical in terms of visual fidelity and things of the sort. That was very very interesting and I was very very impressed overall throughout all these tests with GLM52 being that it is an openw weights model. A lot of folks have been saying just talking a lot about the comparison between these two. Some folks are assuming that Anthropic may be talking more about the dangers of openw weight models and things like this because they're a bit nervous in some of the performance of GLM52. I can't personally speak to that. It's hearsay and it's not based in anything factual. It's just an assessment I've been seeing floating around online a lot. So, I wanted to give a proper head-to-head sideby-side test just on a few different areas of expertise between these two models to see like where GLM52 actually stacks up compared to the current anthropic publicly accessible state-of-the-art. So, in terms of usage here, this is unfortunately where I can't really give a proper comparison. I have only used GLM52 through the CCode app right here for today. And we can see the total token usage. So it could be calculated just based off of this. In terms of the claude code usage, we can just see I haven't really used it at all much this week. So my entire model aotment is 7% used up. The current session which has reset because this video took longer than 5 hours to film is at 13% used. And because both of these are being run through their respective subscription plans is harder to do an independent individual cost breakdown. And my interest in this video was less about price comparison and more about performance comparison. But I did quickly want to touch upon this just because that is definitely going to be a consideration. And the argument could be made that perhaps GLM could be a decent fill-in for alleviating some of the heavier costs of using a state-of-the-art model. Assuming the task one needs it to perform does not necessitate that increasing capability. So, that is going to conclude today's video. I wanted to do a proper head-to-head between these two. This was a lot of fun, especially doing new tests that were not ever used before on the channel. So, I look forward to seeing those propagate through the other channels videos for the next few months. But with that, that's going to wrap it up. If you have any questions, please feel free to leave them in the comments. And thanks for watching.