DeepSeek V4.1 Flash 1017 remade my old animation for under $0.50. Compared with Opus 5.5.
I gave DeepSeek V4.1 Flash 1017 and Claude Opus 5.5 the robot assets from an old voxel-island animation I wasn't happy with and asked each to turn it into a fun, music-driven short. This video contains both final versions: Opus from 0:00 to 0:30, DeepSeek from 0:30 to 1:00. I used DeepSeek V4.1 Flash 1017 through DeepSeek Harness (recorded endpoint alias: deepseek-official / deepseek-flash), and Claude Opus 5.5 through Claude Code. Both initially held on too tightly to the old scene, so I gave them a second brief: keep the robot designs, but freely rebuild the whole environment and art direction. The setting was fantastical floating islands with many agents working, Claude and DeepSeek as the leads, all English, and music driving the animation. I also explicitly told DeepSeek not to economize on tokens. DeepSeek built a floating office, designed its own diving-helmet/whale robot, and wrote a Web Audio soundtrack. Both shorts use Three.js and code-generated music; they aren't outputs from a text-to-video generator. I exported the finished browser animations at 1080p60. I prefer the Opus result visually, but DeepSeek's reported spend was under $0.50, while Opus was around $20 when calculated at API rates. That makes the comparison harder to dismiss. These are my reported/estimated costs, not audited figures, and the different harnesses and feedback mean this isn't a controlled benchmark. What would you ask DeepSeek to improve first: the character movement, the camera work, or the overall art direction? And at that price difference, which result would you choose? Source: github.com/feiyanglyy-stack/animations/.../main/content/films/agent-day-deepseek