$100 AI music videos show sharp gaps between Claude and GPT

A small open-source benchmark put Claude Fable 5 and GPT-5.6 Sol through the same unusually open-ended task: take Bruno Mars and Mark Ronson’s “Uptown Funk”, a lyric transcript, and a fixed budget, then autonomously produce a complete music video.

The setup came from TryAI, which built an agent loop with six tools: plan, web_search, get_budget, generate_image, generate_video, and run_command with ffmpeg/ffprobe. The models could research video generators, choose any FAL or Replicate model, generate footage, inspect clips, edit them, and assemble a final cut. Once the budget ran out, paid generation stopped, but editing could continue.

There were four runs total: each model at $25 and $100. All four completed without hitting time or step limits, and all four produced a valid full-length video muxed with the original song.

How the models spent the budget

At $25, both models nearly used the full allowance. At $100, they did not: GPT-5.6 Sol spent $36.57, while Claude Fable 5 spent $48.60 on generation.

The runs broke down like this:

Recommended reading

EU Forces Google to Open Android AI Access

• Claude Fable 5 · $25: 39m10s, 250,541 output, $24.30, 1280x720
• GPT-5.6 Sol · $25: 42m52s, 386,146 output, $23.18, 1280x720
• GPT-5.6 Sol · $100: 49m39s, 340,702 output, $36.57, 1280x720
• Claude Fable 5 · $100: 38m56s, 280,800 output, $48.60, 1920x1080

The two models made noticeably different choices. Three of the four runs used only text-to-video. The exception was GPT-5.6 Sol at $25, which generated still images first with FLUX schnell ($0.003/img) and then animated them with Wan 2.2-5b i2v ($0.10/s).

At the high end, GPT-5.6 Sol at $100 mixed Wan 2.5 ($0.05/s), Veo 3.1 Lite ($0.10/s), and Hailuo 2.3 Standard ($0.28/video) in a single run. Claude Fable 5 stayed with one model in each run: Wan 2.5 t2v ($0.05/s) at $25, and Seedance 1.0 Pro t2v at $100, priced at roughly $0.12/s at 1080p. Distinct clips per run ranged from 46 to 80.

Token cost and what stood out

Adding model usage changed the totals significantly. Using the published rates cited by the source — Claude Fable 5 at $10 / $50 per 1M input/output tokens and GPT-5.6 Sol at $5 / $30 — the total cost per run came to:

• Fable 5 · $25: $41.29
• Sol · $25: $27.45
• Sol · $100: $39.82
• Fable 5 · $100: $73.65

For Claude Fable 5, token spend alone was $16.99 to $25.05, or roughly 30% to 40% of total cost. GPT-5.6 Sol stayed near $3 to $4 in token cost despite similar token volume.

TryAI’s verdict was mixed. None of the videos were especially good, and all four struggled with character consistency, story coherence, and tempo matching. The systems also interpreted lyrics very literally — including turning “Make a dragon wanna retire, man” into a literal dragon.

The most interesting run, according to the source, was GPT-5.6 Sol at $25, which added text overlays and animated stills with effects that the others did not attempt. But Claude Fable 5's $100 output was the one the authors said they slightly preferred, helped in part by selecting a model with more coherent clips. All four runs used FAL exclusively even though Replicate credentials were available.

The full harness is open source at github.com/hershalb/music-video-arena, and the clearest takeaway is that even on a bounded task with tools, budget, and time to iterate, the models mostly generated clips, stitched them together, and stopped there.