GLM-5.2's Frontend Leap Catches Up to Claude Opus in Blind Tests

Regarding GLM5.2, I have already tested it extensively and written a very long article about it.

My conclusion is: It absolutely cannot beat Opus4.8!

So, saying "GLM5 surpasses Opus4.8" without any qualifiers is a ghost story!

But today, I am not here to criticize GLM5.2; I have already covered its shortcomings.

Today, I am here to discuss an advantage or a point of progress: Frontend!

This morning, upon opening X, the official account of the Large Model Arena posted the following message:

The Chinese translation of this content is:

Exciting news: GLM-5.2 (Max) ranks 2nd in Code Arena: Frontend,

+29 points higher than Claude Opus 4.7 (Thinking), only behind Fable 5!

GLM-5.2 is the best open-source model with the largest lead over Kimi-K2.6 and Minimax-M3.

• React sub-leaderboard 2nd and HTML sub-leaderboard 4th
• Ranked 1st in almost all sub-categories: Branding & Marketing, Reference-based Design, Data & Analytics, Consumer Goods, Gaming, and Simulation.

In one sentence: GLM5.2's frontend has surpassed Opus4.8Thinking, only behind Fable5!

The Large Model Arena is different from pure benchmark data; it involves blind tests with real human participants. Although there is a lot of noise now, with domestic models topping the charts every time they are released, it is still a bit more reliable than benchmarks. While I don't believe this ranking, its ability to reach second place does indicate something.

In our impression, the frontend has been a weak point for the GLM series. Who would have thought it is now becoming a strength!

This is actually good news for those who are really using GLM5.2!

This makes GLM5.2 more well-rounded, and the money spent on it is more worthwhile!

I actually discussed the issue of its enhanced frontend capabilities in two previous articles!

Although I don't know why, it really seems like a refined SKILL has been added. All web designs seem much more polished, and in terms of layout, it surpasses GPT5.5 (mainly because its frontend is too poor).

So, this article focuses on the frontend issue! Don't underestimate the frontend; the most complex part of it is that it directly interacts with end-users, and things that interact with people are actually very complex.

Today, I'll take the opportunity to show you the difference between GLM5.2 and the older GLM5 version.

We can also compare it with Opus4.8.

I won't rant today; I was too tired from testing in the past few days. I'll just quietly do a comparison for everyone. This can help you make an informed decision when choosing a model!

I'll pick a few representative examples.

First, here's a video, and then I'll explain slowly!

1. Cyberpunk "Along the River During the Qingming Festival"

Let's first look at this project, which demonstrates the differences particularly clearly!

The prompt for this task is quite unique, and it's unlikely that major model companies have conducted targeted training for it. So, it is very representative! Only models with strong generalization, aesthetics, and layout logic can do it well. Many models just muddle through!

Let's look at the results directly.

Opus4.8 result:

GLM5.0 result:

GLM5.2 result:

Isn't this comparison very obvious!

GLM5.0 was completely a mess. But with GLM5.2, the shapes of various buildings and objects have emerged. I think it's done quite well. A point for improvement is that it has enough cyberpunk flavor but lacks the Qingming flavor.

Opus4.8's house lines, lanterns, color scheme, and text have a stronger Qingming flavor, and the fusion of cyberpunk and "Along the River During the Qingming Festival" is better.

Of course, there is another contestant that is even more impressive:

Ancient costumes have appeared! There's also the Shangshan Gate and the Bianhe River water level report!

The Rainbow Bridge, an anti-gravity hub! Above are two types of flying vehicles, below are boats and reflections!

Even the distant background scenery has been rendered:

The artistic conception is maxed out!

Alright, alright, let's not talk about it anymore; it's not the main character!

The main character is GLM5.2. We can see a huge improvement compared to the previous version, and it is also outstanding among many domestic models. Because other domestic models are basically at the level of GLM5.0.

2. Infinite Stream Terminal

The Infinite Stream Terminal requires conceiving an adventure story, which can then be advanced through player choices, allowing for endless play, while also considering the butterfly effect. Your choices will manifest cause and effect later on.

This example is inherently a test of intelligence, but unexpectedly, GLM5.2's visual effects were exceptionally excellent! So I'm mentioning it here too!

Opus4.8 result:

GLM5.0 result:

GLM5.2 result:

I won't comment on Opus4.8 for this round; you can go and play with it to see its logical capabilities. We mainly want to look at the difference between GLM5 and 5.2.

GLM5 is in a state of layout collapse, with content overflowing the frame.

GLM 5.2 has a normal layout, and the animation effects are super cool, which you should be able to feel in the video!

Of course, its color scheme and the relationship between light and dark shades are also handled just right!

I even specifically developed a desktop software for this task:

The model used at the time was GLM5.1, and the story was okay. Now, I feel like I can upgrade the visual effects and the model!

3. Gorgeous Gomoku Battle

This task is a dual test of algorithm and UI! The Gomoku project has already appeared in the official ZCode demo, so it should have been optimized.

Let's look at the results directly.

Opus4.8 result is as follows:

GLM5.0 result is as follows:

GLM5.2 result is as follows:

Opus4.8 is a bit too perfect, I won't say more. Its intelligence, aesthetics, and thematic fit are all very good.

GLM5.0's design sense was very poor, and the blue background completely did not match the theme.

GLM5.2 has optimized the board, background, and color scheme. Of course, this theme probably shouldn't have this background color; it could still be optimized.

The gold color scheme helped it recover somewhat. But this example has a Bug... Haha! I analyzed it in the previous article.

4. Neon Runner

This example actually has a bit of a MapleStory feel, constantly pushing forward, supporting double jumps, requiring a map with 3 layers of parallax, and finally, upon death, the character needs to explode.

Opus4.8 result is as follows:

GLM5.0 result is as follows:

GLM5.2 result is as follows:

GLM5.2 is relatively excellent in this round. It grew legs, and there is a slight explosion effect. The key point is that its sense of depth in the scene is very well done, which is a demonstration of spatial capability.

5. 3D Solar System

This example was originally a test of 3D effects. It involves implementing an interactive 3D solar system model using a single HTML file.

Although GLM5.2's 3D effects are insufficient, you can find that its abstract art is played quite well.

Below is the result from Opus4.8:

GLM5.0 result:

GLM5.2 result:

Opus4.8 has maxed out textures, the strongest among all contestants, with the strongest sense of realism.

GLM5.2, on the other hand, played an abstract card this time. It only uses simple lines to construct this 3D solar system. It's not realistic enough, but its line handling is okay, showing some design sense!

All the examples I generated with ZCode are full of "design sense"!

So, it's basically certain that GLM5.2 has undergone specialized training in the frontend. Currently, it seems a bit overly polished, with an excessively strong design sense, which might lead to a feeling that all its works have the same flavor. But it is indeed much better than the previous version, not as rough anymore.

From some of the examples above, it can also be seen that in certain scenarios, the visual and interactive effects are indeed excellent, capable of locally beating Opus4.8. This is good news. Of course, in terms of overall capability, to be honest, the gap is still quite large. There is still room for improvement in processing time, depth of thought, and first-attempt accuracy!

Zhipu's path is correct. It has been looking up to Claude, focusing on improving the model's own capabilities, and it also has high compatibility with Claude. When Opus is not available, GLM5.2 is also a good choice. For top-tier performance, it's not there yet. But returning to real-world scenarios, for domestic users, especially developers, it is still the first choice.

All examples have been uploaded to the topai page of JarvisUni!

This time I also made a small upgrade. In the future, besides the model, I will also note the "agent" tool used to generate the webpage!

The competition now is no longer just about a single model, but an all-around competition. Everyone has started to do engineering optimization and launched their own Harness! So, noting the tool and model will make it more comprehensive!

URL:

https://topai.jarvisuni.com/

The webpage effects can only be seen on the webpage!