Claude Fable 5.2 Generates a 3D Short Film from a Single Prompt, Drawing Attention to the Code-Generated Video Approach
After Anthropic released Claude Fable 5.2, developer Riley Brown gave it just a single prompt, and the model wrote Blender code on its own to complete modeling, camera movement, and animation, then called on 10 GPUs on RunPod for 4K rendering. The process took 52 minutes, cloud rendering cost $5.75, and the final output was a 3D short film titled "The Edge" with 13 shots and a runtime of about 59 seconds. Claude's approach to generating video is to compute and render frames one by one by writing HTML, CSS, JavaScript, Canvas, or Blender Python scripts, which differs from the technical approach of generative video models such as Veo and Runway. The third-party index Claude Video has already cataloged 1,276 videos made by creators using Opus 5.5, involving 1,007 authors, with cumulative views on X of about 78.9 million.
Why read this
AI score 68
The article reviews various cases in which Claude generates video by writing code, and explains how its technical approach differs from that of generative video models such as Veo and Runway.
Topics
Body · AI translation
Yesterday, Anthropic officially released Claude Fable 5.2 and Claude Mythos 5.2. Beyond their impressive programming, scientific research, and robotic manipulation capabilities, Claude's video generation ability also deserves a separate discussion.
After the new models were released, developer Riley Brown posted a 3D short film titled "The Edge" on X.
The visuals revolve around a work desk. The screen of the MacBook on the desk lights up, Claude's lobster mascot Clawd moves around the scene, and the keyboard keys, screen content, and ambient lighting work in concert. The camera keeps cutting, and the sky outside the window shifts from dusk to night, finally greeting the sunrise.
This short film contains 13 shots and runs about 59 seconds.
Riley said that during a livestream he gave Fable 5.2 only a single prompt. The model then wrote Blender code on its own, created the table, laptop, lamp, and character models, arranged the camera movement, and added animation to the objects.
After completing the modeling and animation, Fable called on 10 GPUs on RunPod for 4K rendering, which took 52 minutes. Riley disclosed the cloud rendering cost as $5.75. The model also handled color grading and title creation, and used ElevenLabs to generate the soundtrack and sound effects.
Another developer, Peter Yang, used Fable 5.2 to create a 20-second autumn travel ad for Fushimi Inari Taisha in Kyoto, Japan. The entire video was built through code and includes music.
In fact, back in September, Opus video generation was already a hit
On September 14, developer Kevin Ngo showcased an animation drawn frame by frame by Claude Opus 5 using JavaScript.
Subsequently, a developer used Claude Code to replicate this Risograph stencil-printing style, generating about 3,500 lines of Canvas 2D code, realizing 26 illustrated scenes in a single HTML file, and using WebAudio to synthesize music.
On September 22, Anthropic released Opus 5.5, further fueling this wave of creation.
On the day of release, Reddit user AzorAhai1TK shared a surreal horror short film of about one minute. The request he made to Claude was quite open-ended: make a video about an ordinary workday, with an eerie atmosphere, where both visuals and audio are done through programming, and the model could choose the specific tools itself.
According to the author, Opus 5.5 completed the production in Extra-High reasoning mode in 45 minutes, including the final render. The toolchain involved NumPy, SciPy, Cairo, Piper TTS, and FFmpeg.
In the following days, more complex works appeared one after another.
On September 24, a Reddit user showcased a battle video in the style of a retro SNES game. Opus 5.5 dispatched multiple Agents, respectively responsible for characters, scenes, music, and battle animation. The author said no ready-made assets were provided, but confirmed in the comments that two or three rounds of revisions were made during production.
Around the same period, X user @IterIntellectus posted a 2-minute-16-second animation of the history of Western civilization, linking historical stages such as ancient Greece, Rome, the Renaissance, and the Industrial Revolution with black-and-gold lines and motion graphics.
This video received 5.8 million views about 24 hours after posting, and later accumulated over 15 million plays; the author said only one prompt was used.
The scale of this wave is also considerable. The third-party work index Claude Video currently includes 1,276 videos made by creators using Opus 5.5, involving 1,007 authors, of which 457 come with public prompts. The site counts cumulative X views at about 78.9 million.
How exactly does Claude generate video?
Claude can write HTML, CSS, JavaScript, SVG, or Canvas animations based on natural language requests, and can also use Three.js to build 3D scenes, or call Blender for modeling and rendering.
For example, if a user asks for a 15-second product promo, Claude can first plan the shots and timeline, then write code that controls text, graphics, camera, and transition effects. The program calculates each frame's image according to time, and finally exports an MP4 via browser recording, a rendering engine, or FFmpeg.
If a complex 3D scene is needed, Claude can also write Blender Python scripts to automatically create models, set materials, arrange lighting and cameras, and then call the GPU to complete rendering.
This process mainly relies on the model's programming, spatial reasoning, task planning, and tool-calling capabilities. The visuals are produced through programmatic computation and rendering, which differs markedly in technical approach from the generation methods used by video models such as Veo and Runway.
Code-generated video also has its own advantages.
The positions of text, numbers, and graphics can be precisely controlled; scenes and animation timelines can be revised repeatedly; interface elements in product demos, technical diagrams, and dynamic charts can all be generated using structured code.
Previously, Remotion also launched a Claude Code plugin, allowing developers to create and edit video projects through natural language. On October 8, Anthropic also officially launched the Claude Motion beta, which supports animating text, charts, and graphics and exporting MP4, open to Team and Enterprise users.
At present, the most mature applications of Claude code-generated video are concentrated in motion graphics, information visualization, product showcases, technical popularization, and stylized 3D animation. For highly realistic people, complex natural motion, and continuous cinematic shots, professional generative video models still have their own advantages.
After the release of Fable 5.2, its robotic manipulation capabilities also sparked discussion among several Anthropic researchers. They said on social media that the "bitter lesson" in the robotics field may have arrived.
From Opus 5.5 in September to today's Fable 5.2, Claude video generation has continued to attract attention. Will dedicated video generation models also face their own "bitter lesson" in the future? What do you all think?
This article comes from the WeChat public account "Machine Heart" (ID: almosthuman2014), author: one who follows video generation, published by 36Kr with authorization.