Chinese AI lab Zhipu AI just scored a very public win: its GLM-5.2 model has jumped to the top of DesignArena’s web-design ranking, beating Anthropic’s Fable 5 and pushing past the newer Opus 4.6 and 4.7 variants. The twist is that the model is not only getting better at making polished HTML pages, it is also far cheaper to use than Anthropic’s top-end offering.
That combination matters because web design benchmarks are starting to separate the models that can produce something merely functional from the ones that can produce something people would actually ship. In practice, the winners tend to be the models that understand layout, typography, and visual hierarchy without turning every page into a chaotic demo reel.
GLM-5.2’s rise on DesignArena
DesignArena said on 19 June that GLM-5.2 took first place in its HTML web-design ranking in a single round. The model moved up five positions from GLM-5.1 and now sits at about 1360 Elo in the code category. According to the benchmark, it also improved its win rate by roughly 6 percentage points.
The benchmark’s description reads like a checklist for a competent front-end assistant: cleaner page structures, better image selection from CDNs, tighter typography, stronger visual hierarchy, and restrained animations that add life without turning the site into a carnival ride. It also appears to work well with tools such as Chart.js and Three.js, which is the sort of detail that matters once a model moves beyond static mockups.
GLM-5.2 API pricing versus Fable 5
One revealing detail is the model’s taste in building blocks. GLM-5.2 reportedly uses Tailwind CSS in 91% of its designs and Font Awesome in 51%, while Fable 5 used Tailwind in about 57% of sessions. That may sound like implementation trivia, but it often explains why one model feels more consistent in real projects: it is leaning on familiar parts of the modern web stack instead of improvising every component from scratch.
There is also a price story hiding behind the leaderboard. GLM-5.2’s API costs about $1.40 per million input tokens and $4.40 per million output tokens, versus $10 and $50 for Fable 5. That is the kind of gap that can turn a benchmark win into a procurement decision, especially for teams generating lots of interface drafts, landing pages, or prototype dashboards.
- GLM-5.2 API: about $1.40 per million input tokens and $4.40 per million output tokens
- Fable 5 API: $10 input and $50 output per million tokens
- DesignArena result: first place in HTML web design, category without agents
Anthropic still owns the higher-end narrative
None of this erases what Anthropic’s Fable 5 is built for. The model is still described as one of the strongest systems around, with major gains in reasoning, coding, and vulnerability discovery. But benchmark leadership is fragmented now, and a model can look dominant in one arena while losing ground in another. That is becoming the norm, not the exception.
There is also a political shadow over Anthropic’s AI push. The source material says Donald Trump described the company as a national security threat, then later said he no longer held that view, while access to those models was reportedly cut off for all users after a directive involving foreign nationals. In other words: even the best model on paper can get dragged into policy mess faster than its next update cycle.
The next test is real-world use, not leaderboard bragging
DesignArena rankings are useful because they capture human preference, not just synthetic score-chasing. But the real question is whether GLM-5.2 can hold up once teams ask it to build more than a pretty homepage. If it can keep that design polish while staying this cheap, rival labs may have to answer with better prices, better tooling, or both.