Two Models, One Song, Four Videos

We built an autonomous agent harness to let Claude Fable 5 and GPT-5.6 Sol produce a full music video from scratch. Each model got the same song ("Uptown Funk"), a short text description, a time-stamped lyric transcript, and a hard dollar budget ($25 or $100). They had six tools: plan for thinking, web_search for research, get_budget to check remaining funds, generate_image and generate_video (the only budget-spending tools), and run_command for local shell editing with ffmpeg.

All four runs finished without hitting step or time limits and produced valid full-length videos with the original song muxed in.

The Numbers

RunDurationStepsTool callsFailed callsGeneration spendResolution
Claude Fable 5 $2539m10s2505410$24.301280x720
GPT-5.6 Sol $2542m52s38614610$23.181280x720
GPT-5.6 Sol $10049m39s3407020$36.571280x720
Claude Fable 5 $10038m56s2808000$48.601920x1080

At $25 both models nearly exhausted the budget. At $100, Sol spent $36.57 and Fable spent $48.60 — more budget translated into more footage. Generation spend is the metered FAL cost, not including LLM token costs.

Total Cost Per Run

RunGeneration spendLLM token costTotal cost
Fable 5 $25$24.30$16.99$41.29
Sol $25$23.18$4.27$27.45
Sol $100$36.57$3.25$39.82
Fable 5 $100$48.60$25.05$73.65

Claude Fable 5's token cost ranged $16.99–$25.05, about 30-40% of each run's total. GPT-5.6 Sol's token cost stayed near $3–4 despite similar token volume. Fable uses $10/$50 per million input/output tokens; Sol uses $5/$30.

Tool Choice Divergence

Three of four runs used pure text-to-video. Only GPT-5.6 Sol at $25 used an image-to-video pipeline (generated stills with FLUX schnell, then animated with Wan 2.2-5b i2v). GPT-5.6 Sol at $100 mixed three different video models (Wan 2.5, Veo 3.1 Lite, Hailuo 2.3 Standard). Claude Fable 5 stuck to one model per run (Wan 2.5 t2v at $25, Seedance 1.0 Pro t2v at $100).

Technical Details

  • Prices per second of output video: Wan 2.5 t2v $0.05/s, Veo 3.1 Lite $0.10/s, Hailuo 2.3 Standard $0.28/6s clip, Seedance 1.0 Pro ~$0.12/s at 1080p.
  • Clips generated: 46 to 80 distinct clips per run.
  • FFmpeg usage: All runs used run_command with ffmpeg for beat detection, clip concatenation, and final muxing. Example command from transcripts: ffmpeg -i input.mp4 -filter_complex "showwaves=s=1280x100:mode=line:rate=25" -frames:v 1 output.png.
  • Token usage: Fable 5 $25 used 1,476,900 input tokens and 44,341 output; Sol $25 used 2,956,270 input, 33,220 output plus 9,656 reasoning tokens and 2,558,029 cached input.

What Worked and What Didn't

Character and story consistency: None of the videos held a coherent storyline. Recurring characters drifted between shots.

Literal interpretation: Lyrics like "Make a dragon wanna retire, man" produced actual dragons on screen.

Tempo matching: Cuts landed on the beat (using ffmpeg beat detection), but motion inside clips rarely matched the song's tempo.

Self-review: Models rarely iterated on the edit. Once clips existed, they concatenated and muxed without re-cutting or adding effects. GPT-5.6 Sol $100 shipped some low-quality clips; Claude Fable 5 happened to pick a model with more coherent output.

Budget usage: At $100, neither model spent near the cap. They could have generated consistent character images up front but didn't.

How to Try It Yourself

The harness is open source at github.com/hershalb/music-video-arena. Point it at your own song and budget, swap in models, and see what they build.

Key Takeaways for Developers

  • Agentic tool use varies widely: GPT-5.6 Sol explored multiple video models and image-to-video pipelines; Claude Fable 5 stuck to a single text-to-video model per run.
  • Cost efficiency differs: GPT-5.6 Sol was cheaper overall due to lower LLM token costs, despite generating more clips.
  • Self-review is a gap: Models didn't critically evaluate their own output, leading to inconsistent clip quality.
  • Open-source harness available: You can replicate the experiment and compare models on your own tasks.