DeepSeek's V4.1 Flash scores 81.2 in design tasks, closely trailing OpenAI's GPT-6 Astra while offering a significantly lower cost.
Summary
- DeepSeek's V4.1 Flash scored 81.2 out of 100 in design tasks, achieving 98% of the score of OpenAI's GPT-6 Astra, which scored 82.7, while costing only $0.023 per design compared to Astra's $1.61.
- Among 13 AI models tested—including Claude Fable 5.1 and Grok 4.6—11 performed worse than DeepSeek, with GPT-6 Astra being the only model to score higher.
- The model's architecture uses only 8 billion of its total 552 billion parameters for prompt processing, significantly lowering operational costs.
This week, OpenDesign conducted an evaluation of 13 artificial intelligence models, focusing on their performance in practical design tasks. OpenAI's GPT-6 Astra was the top performer, but DeepSeek's V4.1 Flash closely followed, scoring 81.2 points while charging a fraction of Astra's price.
OpenDesign Arena assesses AI models based on their ability to complete design projects, including web applications and landing pages, using a scoring system out of 100 points. This system emphasizes both design quality and adherence to project requirements.
GPT-6 Astra achieved an average score of 82.7, taking 11.1 minutes to complete designs at a cost of $1.61. In contrast, DeepSeek's V4.1 Flash not only scored 81.2 but also completed designs in just 5.3 minutes at a mere $0.023. Claude Fable 5.1 scored 80.3, costing $3.66 and taking 12.8 minutes.
Of the other models tested, including Grok 4.6 and Qwen 3.8-Max, 11 scored lower than DeepSeek and had higher costs, with GPT-6 Astra being the only one to surpass it by a narrow margin.
DeepSeek's technical documentation reveals that while the model comprises 552 billion parameters, it utilizes only 8 billion for processing prompts and 16 billion for generating outputs. This efficient design is key to its rapid performance.
DeepSeek has previously demonstrated its ability to provide cost-effective solutions, having recently come close to Claude Fable 5 in another benchmark while charging significantly less. The company is also expanding, seeking engineers in Beijing to develop its own operational framework for enhanced control.
The evaluation methodology used by OpenDesign focuses strictly on functional outputs, meaning only complete and operational designs are scored, excluding any that fail to render properly.
GPT-6 Astra, launched on September 3, is recognized for its versatility across various tasks, although some users have noted a decline in its writing capabilities compared to earlier models. Despite its higher cost and slower performance, it remains the top scorer in the evaluation.
DeepSeek V4.1 Flash's delivery rate stands at 57.7%, indicating the proportion of outputs ready for use, while GPT-6 Astra has a delivery rate of 60% and Claude Fable 5.1 has 56.7%.
