Skip to main content

Design Creativity Bench

Today’s AI models can build functional, feature-rich interfaces. But while their technical capabilities have advanced rapidly and been measured across benchmarks, their ability to make creative design decisions remains an open question.

Design Creativity Bench is built to measure this creative ability. It evaluates models across 80 open-ended UI tasks, each adapted to two different product domains. Models are free to make their own layout, structural, and visual decisions, while expert human designers complete the same 160 briefs as a reference.

We evaluate creativity across three dimensions: originality, how distinct a model’s designs are from those of other models; creative range, how much its design approach changes across product domains; and appropriateness, how well the design fulfils the brief.

PaperDatasetGitHub

Creativity Score

GPT-6 Astra
GPT-6.1 Sol
Grok 4.7
Gemini 3.8 Flash
Qwen3.8 Max
Gemini 4 Argon
GPT-5.6 Sol
Claude Opus 5.5
Claude Fable 5.1
Muse Spark 1.3
DeepSeek V4.1 Flash
Grok 4.6
GLM-5.3
Kimi K3
GLM-5.3-Flash
MiniMax M3

Creativity Score vs. Cost ($)

Creativity Score-100102030405060708090100
Human reference 92.3
$0.001$0.003$0.01$0.03$0.10$0.30$1$3$6
Median cost ($) · log scale

AI models fall significantly short of human designers in design creativity

In our benchmarks, GPT-6 Astra leads all evaluated models with an overall creativity score of 65.5, compared with 92.3 for human designers.

Across models working on identical briefs, their originality score sits at 0.592, compared to 0.764 when a human reference was included.

The difference is even more pronounced when the product context changes. Models show limited variation in their approach for two similar domains (clinic vs pharmacy), with a mean cross-pair distinctiveness of 0.581. Human designers adapt their approach more contextually, and this reflects in a score of 0.902.

This gap is evident even in the top-performing model. GPT-6 Astra produces the most original designs compared with other models, yet shows the least variation across paired briefs.

This highlights a fundamental limitation in the creativity of AI models. Models in general explore a narrower range of creative directions, and this limitation is also reflected in their adaptation to different contexts - still far behind from human creatives.

Creative range0.50.60.70.80.90.550.600.650.700.750.80+4.8 SD originality+6.1 SD creative range16 modelsHuman reference
Originality

Models by use case

Best Overall

OpenAI

GPT-6 Astra

Release date: Sep 2026

  • Creativity Score65.5
  • Originality0.646
  • Creative Range0.472
  • Appropriateness99.1%
  • Median Cost ($)$2.564

Today’s models are on par with or better than human designers at following the brief

While AI models fall short of human designers on creativity, the best of them meet 99.2% of a brief’s requirements, a shade above our human designers at 98.0%. All 15 frontier models we tested clear 90%.

When models miss requirements, the errors are typically small and specific. Among the failures we observed, 33% involved missing controls or interface elements explicitly requested in the brief, while 26% involved inconsistencies in generated sample data.

Our results also suggest an interesting trend across model generations. Newer versions of GPT-Sol, Claude Opus, and Grok demonstrate stronger instruction following than their predecessors.

This highlights an important distinction between design execution and design creativity. Models have become remarkably capable of building interfaces that meet a brief’s requirements, but their ability to explore diverse creative directions has not advanced at the same pace. Improving design creativity remains a separate challenge from improving functional accuracy.

Full results

Model
Human reference92.30.7640.90298.0%——
65.50.6460.47299.1%$2.56413.8
63.70.6420.48798.8%$0.46015.8
62.20.6430.60397.8%$0.59318.8
60.60.6370.62897.5%$0.1230.9
58.10.6230.61797.6%$0.30918.0
55.70.5890.61398.6%$0.350—
51.30.5870.50598.2%$0.5984.4
51.00.5570.57499.2%$1.55713.1
48.60.5580.58998.5%$5.32913.9
47.20.5680.56597.8%$0.0801.1
46.20.5800.60596.8%$0.086—
43.60.5760.59696.3%$0.0515.3
43.40.5910.62595.5%$0.1003.5
42.50.5510.63397.0%$0.152—
41.90.5790.63495.6%$0.0432.5
41.50.6750.64391.0%$0.080—
40.40.5470.61596.8%$0.57017.0
38.50.5530.61196.1%$0.02210.0
36.40.6370.69491.2%$0.006—
34.10.6320.63291.2%$0.051—
34.00.5870.60293.4%$0.0242.5
33.20.6620.71189.2%$0.002—
30.10.6100.63991.2%$0.001—
29.90.6650.68388.4%$0.039—
28.10.6270.64689.9%$0.015—
26.00.6490.69488.1%$0.004—
22.50.7300.63183.8%$0.001—
20.60.7380.69682.6%$0.062—
19.00.7230.62283.3%$0.001—
17.70.7380.61782.3%$0.013—
10.60.7360.75380.0%$0.025—
7.10.6940.64081.6%$0.006—
3.30.7010.69680.1%$0.001—
-2.10.8100.70473.7%$0.001—

Questions?
We've got answers.

What does Design Creativity Bench measure?
Design Creativity Bench measures how strongly AI models repeat the same design decisions across different interface-generation tasks. The benchmark measures this tendency by asking models to solve related UI problems where the product context changes while leaving the visual direction and information architecture open to the model. The goal is not to measure whether a design is simply “good” or “bad.” It is to measure how much the model adapts its design decisions when the context should call for a different approach.
How is Design Creativity Bench different from other design benchmarks?
Most existing UI and frontend benchmarks primarily measure a quality or performance metric. For example, Design Arena uses crowdsourced head-to-head preferences to rank generated designs using ratings and win rates, while frontend/WebDev arenas rank models based on how they perform on real-world frontend development tasks. Through the Design Creativity Bench we try to answer a different question: the design decisions made by a model. Rather than asking people to choose which of two outputs they prefer, we deliberately vary the product context while holding the underlying task relationship constant. We then measure whether the model changes its information architecture, composition, and visual decisions accordingly. This distinction matters because capability and creativity are not necessarily the same thing. In our current results, higher model capability does not correspond to lower design repetition; across the evaluated models, the relationship is weak and provides no evidence that more capable models automatically become more diverse designers.
How is the benchmark evaluated?
Models are given open-ended interface prompts where the functional requirements and content are specified, but decisions around layout, information architecture, visual treatment, and composition are left to the model. We construct related prompt pairs that share the same underlying product goal but vary the domain, brand, and content. This lets us test whether the model meaningfully adapts its design rather than falling back to the same default. Each model generates a standalone HTML interface, which is rendered in a real browser. The benchmark then compares the resulting designs using a similarity metric that evaluates their structure and information architecture alongside their colour systems. The metric is separately validated against human judgments and controlled design changes.Read the full methodology
How do you choose which models to evaluate?
We aim to evaluate the strongest publicly available models from leading model providers. Where models expose multiple reasoning configurations, we use their highest or maximum reasoning setting. We prioritize models that perform strongly on established intelligence and UI-generation benchmarks so that the benchmark is measuring differences between already-capable systems rather than simply penalizing models for poor implementation quality. The exact model and reasoning configuration used for every evaluation is published alongside the results.
We have a new model we’d like evaluated. How can we get it included?
We welcome submissions from model labs. If you have a new or upcoming model that you would like evaluated on Design Creativity Bench, reach out to aman@kombai.com with the model details and information on how we can access it. We are also happy to evaluate models prior to release where appropriate.
When was the benchmark last updated?
Last updated: October 2026. Model versions and hosted APIs change over time, so the leaderboard records the models and configurations evaluated at the time of each benchmark run.
Can I see example prompts and model outputs?
Yes. We publish sample prompt pairs alongside the interfaces generated from them so you can inspect both what the model was asked to build and how its design changed—or failed to change, when the product context changed. The benchmark includes examples across different interface types and product verticals, including healthcare, SaaS, retail, fintech, travel, education, and more.
Can I see the generated code and rendered designs?
Yes. Each benchmark generation produces a standalone HTML document which is then rendered in a real browser for evaluation. For published examples, you can inspect the full chain: Prompt → Generated HTML → Rendered interface → Benchmark score This makes it possible to inspect the underlying model output rather than relying only on screenshots or aggregate leaderboard scores.
Can I download the benchmark data and results?
Yes. We make the underlying benchmark materials available so that researchers, model labs, journalists, and developers can inspect the results independently. This includes the leaderboard results, prompt data, supporting evaluation tables, and the code used to reproduce the benchmark methodology.
What are the limitations of the benchmark?
Design Creativity Bench measures a specific dimension of model behavior: how repetitive a model's UI design decisions are under this evaluation protocol. It should not be interpreted as a universal measure of design quality or creativity. There are several important limitations. The current benchmark evaluates desktop interfaces rather than mobile designs; hosted model versions can change over time; and the benchmark evaluates a finite set of tasks and product contexts. We document these limitations, along with the statistical and methodological considerations behind the benchmark, in the paper.Download the paper
Can I report an issue or challenge a model result?
Absolutely. If you believe a model has been evaluated incorrectly, identify an issue in the methodology or data, or would like us to review a particular result, email Aman at aman@kombai.com. We want the benchmark to improve through feedback from model labs, researchers, designers, and the wider evaluation community.
How should I cite Design Creativity Bench?
If you reference the benchmark in research, we recommend citing the accompanying paper: Aman Rusia, Abhijit Bhole, Prashank Gupta, and Dipanjan Dey. “Design Creativity Bench: Measuring creativity in LLM-Generated UI.” arXiv:2610.11539, 2026. You can also link directly to the canonical benchmark page at kombai.com/benchmark/.
Can I use the benchmark charts or results in an article or paper?
Yes. You are welcome to reference, reproduce, and discuss results and charts from Design Creativity Bench with attribution to Kombai's Design Creativity Bench and a link back to the benchmark. For press inquiries, higher-resolution assets, additional data, or questions about a particular result, contact aman@kombai.com or mukund@kombai.com.

Citation

@misc{rusia2026design,
  author    = {Rusia, Aman and Bhole, Abhijit and Gupta, Prashank and Dey, Dipanjan},
  title     = {Design Creativity Bench: Measuring creativity in LLM-Generated UI},
  year      = {2026},
  publisher = {arXiv},
  doi       = {10.48550/ARXIV.2610.11539},
  url       = {https://arxiv.org/abs/2610.11539}
}