Elevator Pitch

• An open-source harness let Claude Fable 5 and GPT-5.6 Sol autonomously create full music videos under fixed budgets, revealing how model tool-use, editing choices, and costs diverge—and where current systems still struggle.

Key Takeaways

• Both models completed four autonomous runs (two budgets each) and produced full-length videos, but differed in pipelines (mostly text-to-video; only Sol $25 used image-to-video).
• More budget produced more footage, yet neither model spent close to the $100 cap; gaps remained in consistency, tempo matching, and self-review/iteration.
• Total cost varied significantly once LLM token costs were included, with Claude Fable 5 notably pricier despite similar outcomes.

Most Memorable Quotes

• “We built a small agentic harness with one job: hand a model a song, a hard dollar budget, and a set of tools, then get out of the way and let it produce a full music video on its own.”
• “Character and story consistency was a struggle for all four.”
• “$100 was probably too much budget. Neither model wanted to spend near the cap, and both kept their step counts modest.”