Claude Finishes 95% of a Model in Blender; Commenters First Ask About Time and Tokens
Yasuhiro Arafune made a test video using Blender × Claude Opus 5.5. Modeling, procedural materials, and BTS all have assigned roles, but public information still lacks prompts, time, and tokens. Look at the visuals first, then the gaps.
Yasuhiro Arafune posted a Blender × Claude Opus 5.5 test video. Here is the footage first.
When AI Fun reposted it, they broke down the approach clearly:
- Modeling: Tripo was given permission, but only about 5% was actually used; the other 95% was completed by Opus 5.5.
- Materials: almost entirely built by connecting procedural Shader nodes.
- Textures: the fine text details on the station sign were handled with ChatGPT's help.
- Sound effects: HunyuanVideo-Foley from Hunyuan.
- BTS: Claude also did the later Making-of with the white model and camera pull.
The most interesting part of this clip is the division of labor. Model generation, procedural materials, text textures, sound effects, and BTS each use their own tools. Reicheru said in a reply that handing everything to AI is less effective than using tools suited to each stage; it is easier to preserve the work's intent and quality. That judgment is more useful than simply shouting that Opus is strong.
But the video doesn't answer the reproducibility question. Nidal Zaiani said directly that without prompts, steps, tools, and time—the actual workflow—the video proves nothing, and a detailed making-of would be more useful. Other commenters asked whether the lighting was HDR and who chose it; whether Claude set the camera movement; how many rounds of prompt iteration there were; and how many tokens were spent. Xiaoguoguo Dunzhou asked even more directly: how long did this run take?
So the 95% figure is shareable, and it also has an information gap. It shows that Opus 5.5 can handle a large amount of modeling and node work in Blender, but it does not mean anyone can reproduce it with the same prompt. Those who want to try should first prepare reference images, storyboards, material-node logic, and check steps. Those who want to question it also have a clear target: wait for Arafune to supplement the workflow.
The comments already have people starting a countdown for traditional modelers. But time, tokens, and prompt rounds have not been disclosed, so don't start the stopwatch on that countdown yet.
Original tweet: https://x.com/arafune/status/2118582485412131107
AI Fun's repost: https://x.com/ai_funss/status/2118621969327198445
发布时间: 2026-11-07 01:26