$100 AI Music Video: Claude Fable 5 vs. GPT-5.6 Sol
Frontier AI models got $100, a song, and full autonomy to research video generators, create clips, and edit a music video—revealing where tool use diverges and creative judgment still fails.
Read Original Summary used for search
TLDR
• GPT-5.6 Sol experimented with multiple video models and creative editing techniques (text overlays, effects) while Claude Fable 5 stuck to one model and simple concatenation
• All four runs struggled with character consistency, took lyrics too literally (actual dragons for "make a dragon retire"), and failed to match clip motion to song tempo
• Neither model iterated or self-reviewed—they generated clips, stitched them together, and shipped without checking quality or re-cutting
• Token costs dominated: Claude Fable 5 spent $17-25 on tokens alone (30-40% of total) while GPT-5.6 Sol stayed at $3-4 despite similar volume
• Both models left budget on the table ($100 cap, spent $37-49) and ignored entire toolsets (never touched Replicate despite having access)
In Detail
The experiment gave Claude Fable 5 and GPT-5.6 Sol complete autonomy to produce music videos: each model got "Uptown Funk," a budget ($25 or $100), and six tools (web search, image/video generation via FAL/Replicate, ffmpeg, budget checker, and a planning tool). No human intervention—models researched available video generators, picked their own parameters, generated clips, and edited the final cut. The entire harness is open source so anyone can reproduce it.
The models diverged significantly in approach. GPT-5.6 Sol at $25 used an image-to-video pipeline (generate stills, then animate), while the $100 run mixed three different video models and added creative editing—text overlays, animated stills with effects. Claude Fable 5 stuck to pure text-to-video with a single model (Hailuo 2.3) and simple concatenation. On token economics, GPT-5.6 Sol was 5-6x cheaper ($3-4 vs. $17-25 per run) despite similar token volumes, making Claude Fable 5 the priciest option at $73.65 total.
All four runs exposed the same creative gaps: characters drifted between shots, lyrics were interpreted literally (actual dragons appear for "make a dragon retire"), and motion inside clips rarely matched the song's tempo despite beat-aligned cuts. Most critically, none of the models iterated—they generated clips, stitched them together, and shipped without reviewing quality or re-cutting. GPT-5.6 Sol's $100 run included genuinely low-quality AI clips it never caught. Neither model used their full budget or explored all available tools (both ignored Replicate entirely). The videos work as technical demonstrations of autonomous tool use, but the lack of taste, self-criticism, and iteration reveals where creative judgment still fails at the frontier.