Claude Opus 5.5 Made a Motion Graphics Showreel From One Prompt. Here's How It Works

By Zubair Hussain · Published September 26, 2026

Key facts

• Anthropic released Claude Opus 5.5 on September 22, 2026. It is available in Claude Code and through the API as claude-opus-5-5 [1][2].
• Within days, creators posted motion graphics showreels they say came from a one-sentence prompt [3][5].
• In one reported run, Claude Code wrote Python for Blender, rendered 900 frames and made the music in about 92 minutes, at about $81 in API costs. Those figures were not independently verified [4].
• The video comes from code that Claude writes and a renderer runs, so every parameter in it stays editable.

AI video usually means typing a prompt into a video model and waiting for pixels. The clips going around this week were made differently. Claude Opus 5.5 writes the code, sets up the scene, runs the renderer and hands back a video file. Anthropic shipped the model on September 22 [1], and by September 25 my feed was full of showreels that people say came from a single sentence [3].

I spent a few evenings going through the posts and the community lists that track them [4][5]. Below is what the model appears to be doing, why the approach matters to developers, and where it falls short.

The prompt that started it

The version that spread furthest was posted by Ajith (@ajith_io) on X [3]:

"make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are, like it's your showreel for a résumé. go all out."

That was the whole prompt. It says nothing about camera moves, colors, fonts, timing, transitions, lighting, 3D objects, render settings or music. Claude was told to act like a great motion designer and left to work out everything else, which is a big part of why the post took off.

Other people ran their own versions. One creator changed the length to 24 seconds, pointed it at their own engineering work, and ran stock Claude Code with Opus 5.5 on max reasoning and memory off [6]. A community list of more than a thousand Opus 5.5 videos groups them by how the frames were made: code-drawn graphics with Canvas, WebGL, Manim or Remotion, 3D work in Blender, edits of existing footage, and pipelines that call outside video models [4].

What Claude appears to have done

1. It wrote the rendering logic

Claude doesn't output an MP4 directly. It writes code that describes what is on screen and how it changes over time, then runs a tool that turns that description into frames. The flow looks roughly like this:

1. Plan the piece: length, scenes, visual style.
2. Build the scenes and objects.
3. Set movement and timing with keyframes and easing.
4. Write the render scripts.
5. Let the renderer output every frame.
6. Join frames and audio into a video with FFmpeg [8].

2. It built a 3D scene in Blender

Several of the showreels are Blender renders [4][5]. Blender has a Python API, so a script can set any object property: position on X, Y and Z, rotation, scale, material, lights and camera position [7]. To animate something, you set a value on one frame, set another value on a later frame, and Blender fills in the frames between them. A minimal version looks like this:

import bpy, math

cube = bpy.data.objects["Cube"]

# frame 1: left side, no rotation
cube.location.x = -5
cube.rotation_euler.z = 0
cube.keyframe_insert("location", frame=1)
cube.keyframe_insert("rotation_euler", frame=1)

# frame 900: right side, one full turn
cube.location.x = 5
cube.rotation_euler.z = math.radians(360)
cube.keyframe_insert("location", frame=900)
cube.keyframe_insert("rotation_euler", frame=900)

The model doesn't have to draw 900 separate images. A few dozen lines describe frame 1 and frame 900, and Blender works out everything in between. A real showreel just has many more objects, lights and camera moves written the same way.

3. It rendered 900 frames

One widely shared run showed a counter reading "900 frames rendered". According to the community tracker, the creator said Claude Code wrote the Python, drove Blender and generated the music, and that the job took 92 minutes and about $81 at API list prices [4]. The tracker notes these numbers were not independently verified.

If those 900 frames make up a 15 second clip, the math is 900 ÷ 15 = 60 frames per second. Film usually runs at 24 fps and a lot of web video at 30, so 60 fps is what you pick for very smooth motion graphics. I'd still treat the exact settings as unconfirmed until the creator shares the project file.

4. It made the music too

The same run lists music among the things Claude produced [4]. Audio is where people use the widest mix of tools: the tracker mentions Node.js scripts and text to speech models for voice, with FFmpeg merging sound and picture [4]. So one agent handled animation, timing, transitions, sound and the final encode, work that usually gets split between a motion designer, a sound designer and whoever exports the file.

How this differs from Sora and Veo

Text to video models such as OpenAI's Sora [9] and Google's Veo [10] generate the pixels themselves. A prompt goes in, the model predicts frames, and a clip comes out.

Claude's route adds a layer. The prompt goes to Claude, Claude writes code, a renderer such as Blender, a WebGL page or Remotion [11] computes the frames, and FFmpeg encodes them. Claude never has to be the renderer. It works more like the engineer operating one.

	Text to video model	Claude writing code
What you get	A finished clip	Source code, assets and a rendered clip
Changing one detail	Generate again and hope the rest holds	Edit a value and render again
Text, logos, prices	Can drift or distort	Exact, because they are strings in code
Realistic people and places	Strong	Weak
Version control	Not really	Yes, the project lives in Git

Why developers should pay attention

Picture an ad made this way, and then a client sends revisions:

• Change the product name and move the logo higher.
• Swap the background and replace the blue objects with red ones.
• Slow the camera rotation.
• Switch from 16:9 to 9:16 and stretch it from 15 to 30 seconds.
• Change the price from $49 to $29.

With a generated clip, most of those mean generating again and hoping the rest stays the same. With code, each one is a variable. Edit it, render, done. The project can sit in Git, go through code review, and be rendered again in CI for every product in a catalog.

The idea I keep coming back to is that the model produces the software that makes the pixels. The scene, camera, coordinates, lighting, easing curves, type, music and render scripts are all files you own, and tools like Blender, WebGL or Canvas do the drawing.

Where code-driven video fits, and where it doesn't

This approach works best for video made of shapes, text, timing and predictable movement:

• SaaS product videos, launch ads and software promos
• Logo animations and YouTube intros
• Animated statistics and data visualization
• Product explainers, UI animations and developer demos
• Portfolio showreels, Instagram Reels and animated presentations

It is a poor fit for realistic footage. If I wanted a person walking through Tokyo in heavy rain with reflections spreading across the street, I'd use a generative video model. If I wanted my SaaS logo animated, three feature cards, a camera move through a 3D space and a finish on the pricing page, I'd ask Claude to build it in code.

How to try the prompt yourself

This setup matches what most creators described:

1. Use Claude Code with Opus 5.5 selected [1].
2. Install Blender and FFmpeg so Claude can call them from the terminal [7][8]. For a browser-based route, Node.js with Remotion works too [11].
3. Run the original prompt word for word first, then ask for changes in follow-up messages.
4. Ask for 30 fps on the first pass. Half the frames means much less render time, and you can move to 60 fps once the design is right.
5. Keep the project folder. The scripts are the part you'll reuse.

Rendering is the slow part. 900 Blender frames on a laptop can take a while, and the API bill grows with every revision loop [4].

My take

The part that excites me is the code underneath the video. A clip that exists as scripts and 3D data can be edited, versioned, automated and plugged into other software. As a full-stack developer, I find that far more useful than one fixed MP4.

For a few years the pattern has been prompt in, content out. These showreels point to a different chain: prompt, then software, then content rendered by that software. That could change how developers, designers and content creators split up the work. The question I now ask of a model is whether it can build the software for the video I have in mind.

Related reading

• Claude Opus 5.5 vs GPT-6: Who Wins?
• 900 frames, 92 minutes, one sentence: the numbers behind the viral showreel

References

1. Anthropic, "Introducing Claude Opus 5.5", Sep 22, 2026.
2. SiliconANGLE, "Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 models", Sep 22, 2026.
3. Ajith (@ajith_io) on X, "Made with Claude Opus 5.5: the 15-second showreel prompt", Sep 2026.
4. athemeroy on GitHub, "awesome-opus-5-5-videos: source-linked guide to 1,000+ Claude Opus 5.5 videos", accessed Sep 26, 2026.
5. opusvideo on GitHub, "awesome-claude-video: curated Claude Opus 5.5 videos, prompts and workflows", accessed Sep 26, 2026.
6. Mr. Buzzoni (@polydao) on X, "24-second showreel variant run on stock Claude Code", Sep 2026.
7. Blender Foundation, "Blender Python API documentation", current.
8. FFmpeg, "FFmpeg documentation", current.
9. OpenAI, "Sora", product page.
10. Google DeepMind, "Veo", product page.
11. Remotion, "Make videos programmatically with React", product page.

---

Written by Zubair Hussain
Full-stack developer (Next.js, React, Node.js) based in Lahore, writing about AI tools that change how software gets built.
Portfolio · GitHub · Blog · thezubairh@gmail.com