LLM-first music: agents wrote the score for a full song for about $50 in tokens

Ok-Bullfrog-3052 · reddit · 2026-11-18

The author demonstrates his 2024 prediction: instead of prompting Suno directly, he had "Claude Fable 5.2" and "GPT-6 Astra" agents co-write an actual score in a shared workspace ($50 in tokens over a week), used Gemini 3.8 Flash for feedback, then had a traditional music model "cover" the result. He specified B minor, chorus dissonance, and bandoneon; rendering was optimized by OpenAI models to run on a 4090 in 9 seconds instead of 44 minutes. He argues prompt-driven models like Suno have hit their compositional ceiling and general intelligence is the future of music.

Original post →

More from Multimodal

Multimodal channel →