Peter Yang Tests Claude Opus 5.5: 3D Scenes, Design, Editing, and Conversation

Peter Yang tests Claude Opus 5.5 across 3D scenes, drawing, app design, video editing and conversational style. He says the model feels more useful and pleasant to talk to than its predecessor, while showing several results he considers imperfect. These are personal demonstrations rather than a systematic benchmark.

Inspired by another creator’s 3D scene, Yang asks Claude to create a Golden Gate Bridge flyover using Blender. He says it works through lighting and key frames; generation takes roughly half an hour and rendering another half hour. He also shows a GPT Astra scene made from the same prompt and finds both impressive.

Next he asks for a browser experience inspired by the Soaring ride, with scenes from around the world. After feedback on the birds, the artifact shows places including the Swiss Alps, the pyramids, Fiji, the Great Wall and Paris, with generated music. Yang says the project took about an hour and likely used substantial tokens.

For a painting app, Yang provides a photo and asks the program to recreate it stroke by stroke in an impressionist style. The result lets him select artistic styles and watch the picture develop. He calls the outcome a promising start rather than a masterpiece.

A second artifact draws images in an anime-like style, first a bridge and then a person with an umbrella. Yang enjoys watching the stages but notices duplicated hair in the finished portrait. He sees creative potential while acknowledging the limits of the examples.

To test computer use, he asks Claude to draw his profile picture using basic shapes in an online Paint-style app. He compares it with a drawing by Astra and dislikes both. This small experiment does not establish a general ranking of browser-use capabilities.

Yang asks Claude Design in Claude Code to propose alternative screens for his fitness app. He finds several generic variations crowded, although he likes one bolder direction. He concludes that a design system and reference images could provide useful constraints.

A more specific request—to simplify a complicated onboarding flow—works better for him. He shows a cleaner sign-in screen and says he can refer to the design canvas when asking for implementation. Yang also says that in his workflow he uses another image generator when he needs a new generated image.

For video editing, Yang improves an existing intro skill using a reference video and applies it to a tutorial. The edit adds captions, zooms, animations and a replacement clip after feedback. He says the agent also blurred sensitive details frame by frame; the result still requires inspection before use.

Yang recommends the open-source HyperFrames project for HTML-based overlays and mentions other editing tools. His demonstration combines a model, an existing skill, an editing framework and his own feedback, rather than showing an autonomous one-step production process.

Finally, he asks an older and a newer model the same self-reflection prompt using his available chats and memory. He prefers the newer answer because he finds it clearer and more actionable, and less judgmental. This is one subjective comparison, not an objective measure of personality or factual accuracy.

Yang’s verdict is positive on creative 3D work, editing and the model’s conversational style, while he wants further improvements to browser and computer use and agent interaction patterns. He invites viewers to try the model in their own workflows.