China’s GLM 5.2 is TERRIFYING!

**Julian Goldie** (0:00)
China's GLM destroys Claude. That's the claim flying around right now. So instead of guessing, I ran the same tests myself. Same prompts, same jobs, GLM 5.2 from China on one side, and then we have Claude Opus 4.8 from America on the other side. And here's what happened. On four out of five build tests, the Chinese model made the better thing, the more fun game, the smoother page. I'm going to show you exactly what we're built today with it and how it works and how this outperforms Claude 4.8 on many of the tests.
So the crazy thing about this is Claude actually lost most of the tests. You might be thinking, now hold on, before you think this is some China good, America bad video, it's not. It's way more useful than that because by the end, you'll know which one to use, when to use it and be smart to move, the beats picking sides and so on. And there's a catch with GLM 5.2 that almost nobody is talking about. So I'll get to that. It changes everything. But first, the tests. Let me set the scene. GLM just came out on June the 13th, 2026 It's built by a Chinese company called ZAI, used to be Jirpu. And Claude Opus 4.8 is built by Anthropic in America, came out May the 28th. Both brand new, both top tier. So I gave them the same five jobs and just watched. The first test was a running game, a simple running game, just dodge blocks, grab coins, speed picks up as you go. And GLM built the fun one. You can see it right here. So this is the test and we use the same prompt on all of these. But this was the most fun one that we actually created. And you can see how cool this is. We actually tested out Kimi KTMOR5, which came nowhere near. But here's the crazy thing. Claude actually lost. I mean, look how boring this game is compared to the other one. GLM 5.2 actually beat it and it created a way more fun situation.
Second test, a landing page.
So we've got GLM 5.2 here. We have Opus 4.8 here. I asked Bo for a clean Apple style launch page for a made up product. GLM 5.2 absolutely nailed it. Let me show you this.
It looked premium. The menu words, the scroll felt nice. Claude's page had less on it and actually felt flat when we tested out. This is Claude's page right here. It just feels a little bit flat, a little bit boring in comparison to GLM 5.2's version. Third test, liquid sloshing in a bowl. And Kimi K217 and Claude Opus made something really boring here. But look at the quality of GLM 5.2. This is GLM 5.2's output. You can change the theme. You can make it more interesting. It just looks way cooler.
Sounds a bit silly to test, but it's a real test of how well the model handles tricky animation. GLM 5.2 made the most fun one. You could change the colors. It feels alive. Claude's actually fades out really fast when you try this out. Seems super boring. Kind of works, but you can only change two colors and they're both really boring. So China won round three. And then there was a fourth test, a neon arcade game. So you can see that right here. Now, if we have a look at GLM 5.2 version, it built something genuinely fun and wild. Look how cool that is, right? That's quite a cool, fun game to play and genuinely fun. Claude's worked, but it feels a little bit buggy when you use it. And also it's just a little bit boring. If you have a look at this, like this is pretty standard, run of the mill stuff compared to GLM 5.2. So China actually won round four. That's four wins for China's AI model, GLM 5.2. Now here's where I have to be straight with you because I'm not here to hype. Claude didn't lose everything on the fifth test, a moving map of the solar system. Claude Opus 4.8 actually made the best version, cleaner, better looking. So it's not a clean sweep. Claude is still excellent at certain jobs, but the pattern is hard to ignore. A model from China that came out this week literally just dropped within the last 24 hours, beat one of the best American models on most of the things that I threw at it. And here's a part that should make you sit up as well. It's the price Claude Opus 4.8 costs $5 per million words in, $25 per million words out. That adds up fast if you're using it all day on the API.