A recent experiment pitted AI models Claude Fable 5 and GPT-5.6 Sol against each other in a challenge to create a music video for Bruno Mars and Mark Ronson's "Uptown Funk" with a budget of $100. An agentic harness provided models with a song, budget, and tools, allowing them to autonomously research video models, generate clips, edit using ffmpeg, and assemble a final cut. Four runs were conducted at $25 and $100 budgets.

All four models successfully produced full-length videos. Spending varied, with models at the $100 budget using more funds than at $25. GPT-5.6 Sol at $25 uniquely employed an image-to-video approach, while its $100 counterpart mixed multiple video models. Claude Fable 5 exclusively used text-to-video methods.

Despite completing the tasks, the videos faced limitations in character consistency, storyline coherence, and tempo matching. GPT-5.6 Sol's $25 run was noted for inventive editing. Claude Fable 5’s $100 run was subjectively preferred but was the most expensive at $73.65 total cost. The experiment highlighted the models' literal interpretations and lack of self-review, with neither utilizing Replicate. The arena harness is open-source for further experimentation.