No Fable. No Sol. Use GLM 5.2!

0:00So, OpenAI just launched GPT 5.6. GPT 5.6 is finally here. GPT 5.6 is here. But you and I can't access it. The in-between state it's in right now kind of sucks. This is really bad. All of us don't get the frontier model right when it drops. I'm I'm so disappointed right now. I really it's hard for me to convey in

0:22words what I'm feeling. Presented by Zapier, the AI automation company. OpenAI is saying that we can't get their latest and greatest model Anthropic had already been scaled back by the government on what they can give us. So, if we can't get the best out of the American companies, I want to learn how to install open-source alternatives, whether they come from China or otherwise. So, today I'm going to watch videos of creators who are going to teach me about GLM 5.2, which I hear is the best. I'm going to understand how to use it, and then I'm going to install it right here, and I've got support from

0:52Adam from Gateway X, the CTO. He's going to watch these videos with me. You'll watch it tutorials with me. We'll understand it, and then we'll go and implement it, and you'll see at the end how easy this can be. Let's get into it. They're really only being beaten by Anthropic's Opus 4.8 and Fable, which is

1:09now no longer available to most people. We can see on some of the benchmarks, they're definitely being beaten by OpenAI as well. Terminal Bench is more of an agentic kind of benchmark there. And the one that probably is most interesting is this deep suite, right? This is a new benchmark that people have been talking about a lot to replace Suite Bench Pro, to replace some of the

1:31ones that have been going on there. What do you think of this? That None of these people even explain why I should care about the benchmarks. Should I care about the benchmarks? You as a regular person probably shouldn't care all that much about the benchmarks. Uh the these benchmarks are there they're all very specific things that are being tested, and each benchmark has its, you know, oh, we're really looking for this, and we're looking for uh biological understanding and health advice or we're looking for software, we're looking for and like most of those things most of the time you don't care about. What you want in a model is one

2:06that solves your problems for you and that's fast and cheap. And the easiest things for any of us like consumers out here picking a model to pick from are the ones that are fast and cheap. You can you can compare token price, you can compare speed. That's quite easy. And then sure, if you're in one that's in the top five, that's great. Also, like try it out. You know, if if if you're not quite happy with what you've with what you've got or this one's a lot cheaper, uh, the easiest thing to do is just take it on a trial for for a day or

2:38for a week. Okay, I have been trying this with my Hermes agent and it's done really well because Hermes agent just chews through tokens cuz it's constantly doing stuff for me and a lot of it is petty. This is doing it and I'm not ripping through my tokens too fast. Okay, here a couple of examples that I think So, it's actually like really solid at design. I want you guys to look right here and see which one of these do you think was designed by GLM 5.2 and which one was designed by Opus. So, as we sort of scroll down here, you can see that we

3:05have, you know, similar style branding and it's obviously the same company. But, we have elements on both that are very similar. We have all of these things come up dynamically as well on either side and there's even a CTA at the bottom. I think the dead giveaway here is this right side was Opus because it has these weird Fs that it loves to

3:22do. It loves that font. Well, you have to I I noticed the exact same thing. I was I'm bummed he gave it away cuz I was like, I think the right side is Opus. [snorts] Uh, so for design, it's really good. By the way, my secret for that is I have my own design and I always give it to Opus and to any tool that I use and say use my design style. Um,

3:41Yeah, that's right. but it did a great job here. Andrew, I like the design elements on the left a little better than the right. GLM I think there's there's uh, this is probably just like a trend thing. So many people have been using Opus for design that it's starting to get a little stale and it and it produces the same stuff over and over again. And just seeing a slightly different perspective here on the left, uh, it's a little fresher. There's a couple of little differences that Opus just wouldn't do by default. And that's one like really positive vote for these open-source models is as there're more and more

4:13open-source models, it's going to be even easier to change and get the opinion of multiple models before you decide something for yourself. Can I say something else? The other way that I was able to tell which one it is is if you look at the tabs on his VS code, it says Opus design design. Okay.

4:30Let's move on to pricing on this though. the fact that you are getting this output for like five times cheaper. So these are the actual terminal sessions where we did those website designs. On the left side with GLM, we got this done in 3 minutes and 59 seconds. On the right side with Opus, we got this done in 14 minutes and 59 seconds. And not only did the left-hand side GLM use less tokens, but its cost per token is also five times cheaper,

4:53ish. That's That's huge for me on a lot of different projects. Are you swayed by this? Are you using this? Will you use it now? I'm interested to try. You know, there's one big downside here that that we haven't talked about yet, which is, you know, when I'm using the Opus models for coding, I'm using my Claude subscription. And so I'm already capped at at, you know, $200 a month for for a lot of usage. And that gives you for $200, it gives you about $10,000 in equivalent API credits. And so it's a little hard, like five times cheaper, that's great, but is it actually cheaper than the

5:32subscription version? That's a very very fair point. Um, I will say this, with a subscription, I still, if I'm working on an intense project, I will run out of usage and they'll they'll going to and I can't stand that. And so, for me, that's where it's an issue. And again, with certain things, it won't you can't use um you can't use Claude. They won't let me put it in into my Hermes agent, for example. All right, let's look at a

5:57coding project. side where I gave them a homework assignment. You can see right here we've got GLM, and then right here we've got Opus 4.8. I had Codex create the homework assignment just so there was no like cross-contamination or anything like that. And then when they finished, I had Codex judge both results and tell us what it thought. Now, in this case, it said that agent two, which was Opus, was better because it handled one subtle edge case that agent one missed, which were duplicate records with values like true versus one or one versus 1.0. So, the short version is that agent one, GLM 5.2, was good, but agent two here was

6:27more precise. And generally, the way that I feel about this so far is that GLM 5.2 is really solid, and it's pretty quick for most tasks that don't require heavy reasoning. Obviously, at the end of the day, Opus 4.8 is a better model. It's a closed-source model. Okay. It does seem like it is better, it does have its place. I have want to start to use it, and we're going to come back to Nate. He's going to show me how I can uh use it in VS Code. Let's take a look though at one thing that you and I have talked about in the past, which is

6:54can we host it locally? Mhm. Here's, by the way, his demo of his local model. So, anyone who's watched my live streams before, by the way, they're coming back soon, you know the 3D first-person shooter test. This what I'm about to show you is the 3D first-person shooter that GLM 5.2 ran completely locally on my Mac Studio. It is a good-looking game. You can see a great environment, good enemies. You can see a lot of video effects. The colors are nice. You can see hit counters and

7:27all that. That's right. Yeah, it is. Um all right. So, now people are saying, "What's great about this is you own it. It's the MIT license. You can do whatever you want with it, put it on your own machine." I've asked you about this. He keeps advocating getting your own machine. hardware you need. You can run local models on any hardware you want. If you have a Mac Mini with 16 GB of memory, there are local models out there that you can run on there. And I'll tell you how to do that a little bit later as well. But for this model specifically, GLM 5.2, it is a beefy model. It is a

8:04chunky boy. You need hardware for it. I am running the two bit quant version of it. We'll talk about that a little bit as well. That version of the model is about 250 GB in size. That means you need 250 GB of memory, which means you can technically run this on a Mac Studio

8:25with 256 GB. Do you agree with that? Um I agree. I think he's he's skipping over something. I've I've tried a number of local models. I've got a a Mac Mini with with like 64 GB of RAM. And so I've tried some of these smaller models that are capable of running on it. The The problem is there's there's a big trade-off between using a model that's big enough to be intelligent, and it fits technically in the RAM, but then it's really slow. And so if you want a fast model, you you're not typically going to be using anywhere close to all of your RAM because you need the extra RAM space for

9:09inference. You need it to run multiple requests at one time. You need it for reasoning and and all of that. So it's oh, well, now I need a much much smaller model so that I get a response in any reasonable time. So it's anytime like for for context here, Andrew, like I when I tap out the RAM on that machine using a model, I will get maybe one token every 5 seconds. That's like the computer is typing one word-ish every 5 seconds. And so writing out a LinkedIn post or coding a file that has 2,000 lines of code, hours and hours and hours for that. Like

9:47totally unreasonable, totally unusable. The only models I've had success with are much, much smaller models that are good for like classification and tagging, but aren't good for coding or intelligence tasks. Let's look at the other option that he gives us. The DGX Station which Nvidia just started releasing across many different providers. That has 750 GB of unified memory, so you can run it pretty easily on there as well. That is just a very

10:15expensive computer. Hold on. I just looked it up while you while you were saying that. Cuz it's like, "Yeah, yeah, that's true." Andrew, it's $85,000. So so okay, so there's a world where locally you can run these models. But yeah, I mean this is if you're spending, this is what I've told you in the past too. If you're spending 60 to $200,000, you can have pretty good performance with these things. Now, I'm not going to put 60 to $200,000 on some machine today. I'm I would much rather pay the the API rates. And especially as API rates for these open-source models come down, why wouldn't you just call the

10:53API? It's so much cheaper. It's so much faster. And then you're not stuck with this like hunk of metal in your home or in your office that may or may not be worth anything in a year. Which one of my past guests had told me broke down on him. Okay, so that is not an option for me. Then the only option

11:09is to come back to Nate. And here it is. If you want to use it, there there two ways that you can use essentially. So anyways, let's talk about how you get set up. What you do is you're going to go to z.ai. And you might pull into something that looks like this. You can go ahead and chat. You can make landing pages right here. You can do 3D modeling. You can build your own mini game. And all of that is kind of free usage because they want you to see what

11:28it does, right? Um I was surprised by how fast and easy it is to get started with it. And it's really good at this stuff. It's really impressive on the front-end design stuff. So, come in here, play around with it if you want to just test it out and see how it just say for anyone who's listening, just go to their website. It's z.ai and just start asking it to do dumb stuff to see what happens and I think you'll be

11:48impressed by by what you get. Yeah. But then if you want to plug it into Cloud Code or Open Code or Hermes Agent or wherever you want to plug it in, you're going to click on this button up in the top, right? And that's going to take you to the actual API console. And so, you've got the option to just pay per token, which obviously is not too expensive. If you come here and I go to my billing and I go to the model pricing, you can see on the input it's $1.4 and on the output it is where to go

12:11$4.4. But you can also just get a plan. So, you could go on 16 bucks a month, 64 bucks a month, or 144 bucks a month. And if you go yearly, you can save even more money. Obviously, they're not a sponsor. I'm not working with them, but you can also get on a plan. So, maybe you can be on a Cloud plan for maybe 100 bucks a month and then you can be on a Z plan

12:29for 64 bucks. That's my plan. What do you think? I mean, that that makes a lot of sense. Do you get API access on the plans or is it just access to the user interface? Yes, you get API access on the plans. That's like that's banked up. Like you get some you get a bunch of credits

12:47every month. Yep. So, I I've gotten out the light plan at I only did monthly cuz I said who knows what's going to come out next month. $16.20 a month. I've got it in my Hermes Agent. I tried it to see, Adam, is it reasonable? Can I talk to it like a human being the way that I could with Claude and if frankly even with Codex, everything's been working great. It even will code up pages like the one you're

13:08looking at now for me. Um and so, I figure, okay, this is good. Now I want to put it into VS Code. But so far, I I see you liking this, right? Yeah, that's amazing. I I didn't realize the plans also cover it. That's I mean, that now now we're like getting cherry on top category, you know? Can we get it set up in your in your VS code and and

13:28and make something? So inside of your dot cloud, you should have a settings.local.json. And this is where you can play with your permissions and your MCP servers and all that kind of stuff. And you can also set environment variables. So if you guys have activated agent teams, this might live here like mine, or it might live globally. Wherever it lives, that's an environment variable. And all you need to do is you need to come in here and you need to set these. So what I'm going to do is I'm going to have this exact thing pasted in the description of this video. So all you have to do is copy

13:54this, put it into your environment variables. You could also say, "Hey Cloud Code, put this into my settings.local.json." And then all you have to do is switch out your own API key. So this is where you will put your Z API key that you just got in here when you go to API keys and you added one right there. I think

14:09we're ready to do it, right? I think we're ready. Let's let's grab the screenshot of what Because what you're doing here as you can see, you have this Do a screen capture. Actually, let's go back and make sure that we get it from here so that it's complete including the the place where it gets stored. And I'm going to stop sharing this. I will switch now to VS Code where I will open up a brand new experimental thing. So here we go. This

14:37is my Cloud Code right now. I'm going to get rid of this. I'm going to paste the screenshot and say I want to use Z.ai's GLM 5.2 model here. Take a look at this screenshot. In a moment I'll give you my API key. Tell me what I need to do. There it is. Just type that in.

14:59All right. While I'm doing [snorts] that, I'll get an API key which I'm going to go to the same place that he told us to go. Okay. And this is what I need from you, the API key, which I'll put right here. I'll need to restart the Cloud Code in this folder, okay? And then let's plug it in

15:15and see.

15:22Do you think it's going to work? Neat does seem to give good information, so I'm going to say yes, it will. Okay. I think it'll work, too. [music] Okay, what model are you using? I'm Claude Code, uh but I'm running Okay. Great. That's unusual. [laughter] Uh the folder name gives it away. Okay. Uh how do I switch

15:51back? One one easy way to actually take a file like this and and sort of delete it is to just to change the name. Okay. So, if you wanted three different models in here, I'd have like settings.local.json, and then I would add like .zai at the end when I wasn't using that file. Got it. You know, and so it's like Anthropic's going to or Claude Code's going to see whichever one is called settings.local.json, but you could have three or four

16:17stand-ins ready to swap out with it. So, this was pretty freaking easy. This feels like the answer. Pay for an extra z.ai account that I use in Hermes in places where I can't use Claude Code, have it as a backup so that I could come in here and switch whenever I run out of credits or if I have something that's really intense, I just switch to this and let it do the intense work. This is fantastic. And Adam, I'm a whiz when it comes to this stuff. I don't want to mess around with anything that's called settings.local.anything, and now I was

16:45able to do it easily. Yeah, I mean, you can ask Claude to to help you with it. I I think that's the like the cheat code that most people aren't using is people feel stuck, they feel afraid, I don't know what to do. Just ask. This thing is sitting here right here ready to help, already on your screen. And you know, other people will try and create the file and paste the stuff in. You did it exactly how I would do it, Andrew, which is I give Claude the information that I have and my request, and I let Claude handle making the files, make the changes, set

17:15up the backup, all that stuff. Like it's it's a going to be a lot better at that than than you are. And And honestly, it's it's going to be faster, too. All right. And one of the best things about this type of video is that you get ideas for what you can do. The how do you act on it? Give it to Claude. It'll do it. The what can you do? You need ideas for, which is why I've got your next video here about my favorite agent, which is Hermes. You're going to get a ton of ideas for what you could do with it. And now that you've got an

17:39inexpensive brain for it, you're going to be able to do a lot more. The video's right here. Click and follow. See you there.