How GPT-5.6 and Kimi K3 Learned Good Design — Design Arena Research

Design Arena, a blind comparison benchmark where humans vote on AI-generated websites, released analyses of two top models: GPT-5.6 Sol (OpenAI) and Kimi K3 (Moonshot AI). Sol achieved an Elo of 1353 on Web Design (Non-Agentic), surpassing its predecessor GPT-5.5 by 18 positions, and set new Pareto frontiers for speed and cost. Researchers found that Sol learned to suppress AI antipatterns (e.g., purple gradients, bento grids) by refusing to generate them, creating "holes" in its design space. In contrast, GLM 5.2 avoids antipatterns by not having learned them at all. Kimi K3 topped single-shot Frontend Arena with Elo 1408 and 3D Design Arena with Elo 1450. K3 simulates a full AI agent inside its reasoning chain, writing 10 times more code in thoughts than other Kimi models, and uses a learned "internet index" to recall real image IDs from Unsplash without broken links. Each model employs a distinct strategy: GLM 5.2 avoids bad design by ignorance, Sol by active suppression, and K3 by iterative mental simulation.