DeepSeek's New Model Nearly Matches GPT-6 Astra on Design—at 1.4% of the Cost
OpenDesign Arena evaluated 13 AI models on practical web-design tasks and found OpenAI's GPT-6 Astra scored highest at 82.7 out of 100. DeepSeek's new V4.1 Flash trailed by just 1.5 points while costing roughly 1.4% of Astra's per-design price, completing tasks faster and using a technical design that activates only small subsets of its full parameter set.
Why It Matters
The results highlight a growing trade-off in model development: similar real-world design quality can be achieved at a small fraction of cost by changing model architecture and inference behavior. For teams choosing tools for everyday design work, price and completion speed may now weigh as heavily as top-line benchmark scores.
Key Facts
- benchmark: OpenDesign Arena tested 13 AI models on real-world design tasks
- top-score: GPT-6 Astra scored 82.7 out of 100
- deepseek-score: DeepSeek V4.1 Flash scored 81.2 out of 100 (98% of Astra)
- models-behind-deepseek: 11 of the 13 models tested scored lower than DeepSeek V4.1 Flash
- cost-per-finished-design-astra: $1.61
OpenDesign Arena ran a head-to-head evaluation of 13 AI models on tasks typical for working designers—building web apps, dashboards, mobile screens and landing pages—and graded outputs on a 100-point scale. The scoring splits 30 points for meeting the brief and 70 points for design quality, including layout, hierarchy, color and style fit. Only outputs that rendered as a working webpage were scored; blank, broken or cut-off results received a zero and were not retested.
OpenAI’s GPT-6 Astra topped the chart with an average of 82.7 points, taking about 11.1 minutes and $1.61 to produce a finished design. DeepSeek’s V4.1 Flash finished a close second at 81.2 points, completing work faster (5.3 minutes) and far more cheaply—about $0.023 per finished design, roughly 1.4% of Astra’s cost. Of the other models tested, including Claude Fable 5.1, Grok 4.6 and Qwen 3.8-Max, 11 scored lower than DeepSeek while also charging more to run.
DeepSeek attributes its low cost and speed to a Causal Encoder-Decoder architecture that keeps its full parameter count dormant for many prompts. According to DeepSeek’s technical paper, the model contains 552 billion parameters but only activates 8 billion to read an incoming prompt and 16 billion to generate the response. That selective activation is the company’s stated explanation for both the model’s quick completion times and the reduced inference expense.
OpenDesign’s delivery-rate metric—the share of outputs judged ready to hand off without revision—was similar among the top models: GPT-6 Astra had a 60% delivery rate, DeepSeek V4.1 Flash 57.7%, and Claude Fable 5.1 56.7%. The report notes GPT-6 Astra’s reputation as a generalist capable of varied tasks and that some early testers found its writing weaker than its predecessor. DeepSeek has been iterating rapidly: its prior V4 Pro model previously closed most of the gap with Claude Fable 5 on a different benchmark while charging much less, and the company is recruiting engineers in Beijing to build a Code Harness to expand its stack beyond the model itself.
OpenDesign’s setup narrows what these numbers prove: the benchmark measures reliable, everyday design output rather than broader reasoning or coding ability. For design teams weighing tools for production work, the results suggest near-parity in output quality is achievable at dramatically different price and speed points depending on model architecture and inference strategy.
Keep Reading
This Guy Cloned Sam Altman, Elon Musk, and Zuckerberg Into AI Bots. They Immediately Started Fighting
Apple is walking a ‘tightrope’ when it comes to the prices of its new iPhones
Panic builds over bankrupt Spirit’s looming data sale to Google
