October 9, 2026
A music video you can edit piece by piece: GPT-6 Astra Ultra coded the music and Blender scenes
In an Oct 9 experiment, an AI agent created the two-minute music video Leave a Way and saved the music and scene source files.
A folding ruler becomes a bridge for a stuck bead. The agent built this story in Blender and wrote a Eurodance song for it.
Source files beyond the video. The creator, mika314, required editable deliverables and verifiable constraints from the outset. According to his report, the agent worked for about 4 hours from a single 2,778-word prompt, with no further instructions. Alongside the video, it created Blender scenes and MIDI files, saved separate audio tracks and kept the source scripts.
The agent composed the music at 128 BPM using MIDI, synthesis and audio processing, with 7 samples. Vocals came from a local Hoshino Hanami/DiffSinger setup, while Gemma 3 1B and Llama 3 8B helped brainstorm ideas. The creator used existing tools and asset libraries, without Suno, Midjourney or Sora.
One run, several revisions. During the run, the agent removed the drums and bass for 6 bars as the protagonist made a decision. It fixed intersecting geometry, set up a portrait camera and synchronized the text's appearance with the sung sounds. A single initial prompt here meant working without the creator's intervention, rather than making just one attempt.
The agent wrote checks for note pitch and vocal timing, audio clipping and loudness. It separately checked the audio mix assembled from the tracks, text placement in the frame and gaps between objects, as well as video decoding.
Both exports were 120 seconds long, with 3,600 frames each and the same audio track.