China’s GLM 5.2 is OUTRAGEOUS!

**Julian Goldie** (0:00)
China's Glm 5.2 is absolutely outrageous. And I'm going to show you some of the stuff we've built with it. You can see some examples of really cool things. This is frontier level intelligence at a level I've never seen before when it comes to Chinese AI. Here's another example. So this is a full open world game we've created. Number one, it looks pretty amazing in terms of the colors and everything. And number two, this was just coded out with like just one single simple prompt. It can build amazing things as you can see right here. The graphics are super nice. Here's another thing that we built out. Here's another one. There you go. Here's another one as well. So if we play this, you can see like how wow this is, right? Like it's a lot of fun to build these things out. And it can just automate anything. And this is not just for games, just like kind of visual examples to show you. But you can automate whole worlds. You can, for example, automate tools, games, websites, apps. And this is all things that we've just built in a single morning since the benchmarks came out. So it's absolutely shocking what you can build with this and how you can build with it and also how easy it is. So I'll show you exactly how to use it in a second. Here's another example.
You can build like full 3D games, et cetera. And the thing I've seen about this as well is it's actually outperforming on many benchmarks, GPD 5.5. It absolutely dominates and crushes Gemini 3.1 Pro on the benchmarks. It's way cheaper. And it's right up there with something like Claude Opus 4.8. Now you may say as well, OK, it's not as good as 4.8. I would agree with you. I'm 100% with you if you believe that. But at the same time, bear in mind, like for the Claude Coning plan, we're paying like $200 a month. Whereas, for example, this to build all of this out was like $80.
And I've not even come close to hitting the token limit. Plus, one of the biggest differences here is that we can actually plug Glm 5.2 into our Hermes agents. We've got Hermes Jarvis here. It's a voice activated version of Hermes. And basically, we can control it, and I can plug in my coding plan with Glm 5.7 to automate whatever I want with it. Whereas, for example, if you're using something like Claude, you can't plug it into Hermes. Good luck using an AI agent, you can't, right? Which means you're a massive disadvantage. So if you've never used this before, if you don't know what Glm 5.2 is, or you just want to see how to build with it, this guide is for you. I'm going to talk you through everything. And let's kick it off. So Glm 5.2 is from ZAI. ZAI is Jirapu. It's quite a small AI lab in China. They come from Beijing. And you can see an example of the benchmarks right here. Now, they were basically non-existent about UO. No one had even heard of them. And now you can see they're right up there with Glm 5.1. If I actually open up my Twitter thread, it has been bombarded with announcements about Glm 5.2 today. So it's pretty wild. And if you look at the benchmarks here, you can see for Frontier SWE, so Long Horizon Task Completion, which is, you know, where you can give agents a task and they just go off of it for hours. Or you can see here that for Frontier SWE, Claude Opus, of course, killing it, killing it, right? But Glm 5.2, 74.4. They are so close. You might say, well, benchmarks, we don't care about those. Benchmarks, they don't matter. I would agree with you. That's why I built out all that cool stuff before, just to test out for myself and see it was actually good. And it was actually good. And you can see here, for example, if we compare this versus Glm 5.2, it's within 1%.
74.4 versus 75.1. If we look at GPD 5.5, not even in the race, really. And where is Gemini 3.1 Pro? It's not in the race at all. It's not even up there, right? It's not even comparable on the benchmarks. And I think this is pretty shocking for Google because they've got to bounce back now with 3.5 Pro. That's got to come out soon. Otherwise, they are lagging behind all three of these models. And bear in mind, GPD 5.5, totally closed. Claude Opus 4.8, totally closed. Glm 5.2, open source, my friends, open source. That means that if you have a good setup, you can just go and run this locally for free forever. How insane is that? How insane is that? And you can see some of the cool stuff we've built with it. Like, this is like a full open world game here. It's not like we're just building trash. You know, this is actually like pretty interesting stuff that we've built. And this is without even trying. You know, we just spent a couple of hours. It's not like we've spent months creating this. You can also, for example, create videos, as you can see right here. And because we've got Hermes working with this, we can have a team of Hermes agents. They just chip away and create an amazing avatar like you can see. And then they go off and automate this fully for us. It's wild. This is wild stuff. I've never seen this before. Now you might be saying, okay, does it beat Claude Opus 4.8 or anything? You can see it's actually ranked number one on Design Arena first on Design Arena, right? It's got an ELO of 1360, which means that it is jumped ahead on this shocking. This actually shocks me. And whether this is like, again, try it for yourself. And I've tried it for myself. I really like it. But you can see here that it has jumped ahead of Claude Fable 5