The OpenDesign Arena benchmark evaluated 13 AI models on design tasks including web apps, dashboards, and landing pages. OpenAI's GPT-6 Astra achieved the highest score of 82.7 points, while DeepSeek V4.1 Flash followed closely with 81.2 points.
The cost difference between the two models is significant. Generating a design with DeepSeek V4.1 Flash costs $0.023, which is approximately 1.4% of the $1.61 price tag associated with GPT-6 Astra. DeepSeek also completed tasks faster, taking 5.3 minutes compared to 11.1 minutes for the OpenAI model.
Out of the 13 models tested, including Claude Fable 5.1, Grok 4.6, and Qwen 3.8-Max, 11 models performed worse and cost more to operate than the DeepSeek V4.1 Flash. The benchmark measures output reliability, requiring designs to render as functional webpages to receive a score.
DeepSeek V4.1 Flash utilizes a Causal Encoder-Decoder architecture to achieve efficiency. While the model contains 552 billion parameters in total, it activates only 8 billion parameters to process a prompt and 16 billion to generate a response.
The delivery rate, representing the percentage of outputs ready for use without revision, reached 60% for GPT-6 Astra and 57.7% for DeepSeek V4.1 Flash. DeepSeek is currently expanding its engineering team in Beijing to develop its own agentic stack.
