On November 20, 2026, Artificial Analysis compared GPT-6.2 Astra and Dreamina Seedance 2.5 across three scenes using identical camera paths: the Sydney Opera House, an exploded mechanical watch, and a dragon attacking a medieval town. Astra did not generate video directly, but produced Three.js code that allowed the scenes to be explored interactively in a browser at 1080p and 24 frames per second. Seedance 2.5 generated native videos from prompts and reference images, supporting clips of up to 30 seconds and as many as 50 reference assets per request. Its API costs roughly $0.23 per video second, while code-generated scenes are produced once and then rendered on the viewer’s device. The test highlights the complementary strengths of models focused on three-dimensional scene construction and dedicated video models specializing in visual rendering, texture, lighting, and atmosphere.
Source: Read the original article


