OpenDesign Arena This week, the same set of design task tests were conducted on 13 AI models. The results show that DeepSeek V4 .1 Flash performs close to OpenAI's GPT-6 Astra in real design scenarios, yet its operating costs are significantly lower than the latter.
The test results are close to those of the top model.
OpenDesign Arena mainly evaluates daily design tasks such as web applications, dashboards, mobile interfaces, and landing pages. Out of the total score it assigns, 30 points are based on whether the task is truly completed, while 70 points are determined by the matching of layout, hierarchy, color scheme, and style.
In this test, GPT-6 and Astra ranked first with 82.7 points. DeepSeek V4 scored 81.2 points, which is 98% of the former's score. The difference between the two is only 1.5 points.
There is a significant cost difference.
The data provided by OpenDesign also shows that it takes an average of 11.1 minutes and costs $1.61 to complete one design with GPT-6 and Astra. With Flash, the average time taken is 5.3 minutes, and the cost per instance is only $0.023.
Among the 13 models, except for GPT-6 and Astra, the scores of the remaining 11 models are all lower than DeepSeek V4 by more than 0.1, and their operating costs are also higher. The models that participated in the test also include Claude Fable with a score of 5.1, Grok with 4.6, Qwen with 3.8, Max, Kimi K3, GLM-5.3, Flash, and Gemini with 3.8, as well as Flash.
Low prices come from the parameter activation method.
According to the technical report of DeepSeek, the total number of parameters for V4.1 Flash is 552 billion, but only 8 billion parameters are activated when reading prompts, and 16 billion parameters are activated when generating responses. The company refers to this design as Causal Encoder - Decoder.
OpenDesign also reminds that this type of test is more focused on measuring the ability to stably output web pages, rather than general reasoning or programming skills. GPT-6 Astra is a new model released by OpenAI on September 3rd, and it still maintains a leading position in this design ranking.











