Our benchmark reports a combination of three evaluation metrics: expert pairwise preference, the Design Critique Rubric, and an AI Slop Detector. The leaderboard ranks the models by expert preference. This page covers the other two.
Design Critique Rubric
Just an Elo score doesn’t tell you anything about how to improve the model. We acknowledge that there are design principles all good designs follow.
In Frontier Design, we curated a list of 35 high quality design rubrics based on widely acknowledged design principles. These design rubrics capture common design mistakes such as using too many fonts or spacing inconsistencies. Grading the design outputs on these rubrics can establish a baseline of design without mistakes. The rubric wheel groups them into three areas: structure, visual craft, and content and use.
The rubric was developed based on designers’ critique on the Human Creativity Benchmark. We tested the rubric’s judgments against what designers say about the same pages. Where designers had left written critiques, we compared the issues the rubric raises with the issues designers raise. Designers also reviewed the rubric’s critiques one by one, with decoy critiques mixed in.
The design rubric ranks the models much the way designers do. In our analysis, the results of our design rubrics against AI model outputs have a high correlation with the pairwise ranking of models (r = 0.94, Spearman 0.95).
Slop Detector / AI Slop Patterns
As part of the Frontier Design benchmark, our research team used statistical analysis on AI model outputs to create an internal Slop Detector engine that identifies AI model signs and tells. We supplemented this engine with a hand-curated AI slop pattern list that our design team has curated over time. As a result, when using our Slop Detector to score a website, we can receive an AI-likelihood score, as well as all design details that support the score.
The meter reads a few hundred features of a design, in 18 dimensions, and indicates how close the design is to AI-made websites. The AI slop patterns sit in four of the visual dimensions, marked below.
Visual
- Colour and surfaces19
- Typography24
- Labels, numbering and emphasis24
- Repeated components23
- Shapes, cards and decoration28
- Layout29
- Data displays27
Content and writing
- Audience20
- Value25
- Offer17
- Explanation29
- Proof27
- Opening30
- Section order25
- Navigation33
- Calls to action31
- Copy26
- Voice27
Holds AI slop patternsNumbers are features per dimension

The Slop Detector tells human work apart from model work, and it picks up each model’s quirks. On a 0 to 100 scale, the human golden designs read 0.9 and the models read from 1.2 to 66.
- Human golden
- Gestalt
- DeepSeek V4 Pro
- Kimi K3
- GLM-5.3 Prime
- Grok 4.7
- Claude Fable 5.2
- Claude Opus 5.5
- Qwen3.8 Max
- Muse Wonder
- Muse Spark 1.3
- Claude Sonnet 5.5
- GPT-6 Astra
- GPT-6.1 Sol
- Gemini 4 Carbon
- Gemini 4 Argon
- Kimi K3.1 Pro
DeepSeek V4 Pro reads as low as the human reference because many of its pages are unfinished: they look unlike a typical AI-made page, not more like a human-made one.