Agent Builds a 4K Blender Video and Renders It on 10 GPUs in 51 Minutes for $5.75

A 3D ad-style video usually takes days: modeling, animation, lighting, a render farm, sound. The Telegram channel "Tochki nad II" ("Dots over AI") shared a video that was built entirely by an agent running Fable 5.2 from a single prompt. The result: 4K, 51 minutes of rendering and $5.75 in costs. Here is what actually happened and what you can take into your own work.

What exactly the agent built

The prompt asked for little: create a room in Blender with a laptop where vibe-coding things happen, as in an Anthropic ad, show the Clawd mascot and render everything on RunPod. After that the agent acted on its own.

First it wrote a Python script that built the whole scene: the room, a MacBook with 77 individual keys and Clawd himself. Blender is controlled through Python, so for the agent this is an ordinary coding task: write the code, run it, look at the result.

The nicest detail is the screen. The model built a web page itself, saved it frame by frame as 1345 images and mapped them onto the laptop lid. The keys depress exactly in the frame where the matching letter appears on screen. In effect, the agent synchronized two layers: the content on the screen and the physics of the keyboard. It is exactly such small things that separate careful work from "good enough."

The numbers and how the pipeline works

Then came the scale: 13 cameras, a simulated sun from sunset to sunrise and a 4K render on ten rented graphics cards. It all took 51 minutes and cost $5.75. The model also generated the music and sounds in ElevenLabs.

Note the structure. The agent does not "draw video" with a neural network; it orchestrates ordinary tools: code builds the scene, web layout supplies the screen content, cloud GPUs split the heavy render among themselves, and a separate service produces the sound. This approach is fundamentally different from generative video. Every element of the scene remains an object in a script that can be edited: change the wall color, remove a camera or rewrite the on-screen text, then restart the render without rebuilding everything.

For a product manager or founder the economics matter here. A video with 13 angles and a 4K render costs a few dollars of GPU rental and about an hour of waiting. When production gets this cheap, the bottleneck becomes the idea and taste, not the render.

Limits and practical takeaway

Now, honestly, about the limits. This is a single Telegram post, not a reproducible experiment. It is unknown how many attempts and fixes were needed, how the agent verified the result, and whether the $5.75 includes the cost of the model's tokens. Most likely a successful run is shown, not an average one. The post gives no technical details about Fable 5.2, and at the end the author himself asks whether anyone has tried running Blender through an agent on the new model. So repeatability is not yet confirmed.

What to do with this in practice. First, look at your own programmable tools: many 3D, CAD, video and graphics editors have a scripting API, which means they can be handed to an agent as a working environment. Start with a small scene of three or four objects and one camera to understand where the agent makes mistakes. Second, plan for cloud rendering from the start: renting several GPUs in parallel for an hour cost just a few dollars in this example. Third, ask the agent to save intermediate results and scripts to a repository so the scene can be edited and versioned like ordinary code.

The main signal is not the $5.75 but the type of task itself. The agent carried a long chain of code, layout, cloud compute and sound through to a result without requiring manual human work in between. For now this is a single demonstration, but these are exactly the cases worth repeating on your own data before planning a product around them.