$100 AI Music Video: Claude Fable 5 vs. GPT-5.6 Sol

TL;DR Highlight

Given a budget and tools, two LLM agents made music videos; Claude Fable 5 cost up to $73.65 total.

Who Should Read

A backend developer designing long-running agent loops who wants to know how tool use and cost differ across models.

Core Mechanics

Six tools in the agent loop
- The agent got plan, web_search, get_budget, generate_image, generate_video, and run_command (ffmpeg).
- Only the two generate tools spend budget; at zero, generation is refused but editing continues.
All four runs finished
- Claude Fable 5 and GPT-5.6 Sol each ran at $25 and $100, four runs total.
- None hit a step or time limit, and all produced full-length videos with the original song.
Models chose different pipelines
- Three of four runs used text-to-video only; only Sol at $25 made stills first and then animated them.
- Sol at $100 mixed Wan 2.5, Veo 3.1 Lite, and Hailuo 2.3 Standard, while Fable at $100 used only Seedance 1.0 Pro.
Fable 5 had higher token costs
- Total cost was $41.29 (Fable $25), $27.45 (Sol $25), $39.82 (Sol $100), and $73.65 (Fable $100).
- Fable's tokens cost $16.99 to $25.05, about 30-40% of total, while Sol's stayed around $3-4.

Community Reactions

Videos copy lyrics too literally, commenters said
- One commenter noted good music videos use a story arc rather than literal lyrics.
- Another cited 'don't believe me just watch' showing only an arm wearing a watch.
Human involvement does much better, per commenters
- One said Kling clips were not one-shot and were finished in a video editor.
- A human-in-the-loop workflow example was shared.
Older video models were used, one commenter said
- They said the models used year-old video models capped at 10-second clips.
- Seedance 2 reaches 15 seconds, 2.5 reaches 30, and both support multimodal references.
Agents should iterate and review, one commenter argued
- Filmmaking is iterative, so agents should review at every step.
- They also estimated the original video's budget at about $100,000, calling this a 100x cost cut.
Taste can't be offloaded, a commenter said

How to Apply

Add a budget tool to your agent
- Today, add get_budget to a long-running agent loop and deduct only for paid tools.
- At zero, refuse paid calls but keep local editing tools allowed.
Compare models on token cost too
- When running one task on two models, log generation cost and LLM token cost separately and sum them.
- Pick models by total cost.
Add a self-review step to the loop
- Require a check right after generation in the system prompt.
- Tell the agent to regenerate or re-edit when output misses your criteria.
Make reference images first
- For tasks needing consistency, have the agent create a reference image first and reuse it.
- The source did not try this, so verify the effect yourself.

Try It

copy & run

# Paste into Claude Code or ChatGPT (simplified version of the source experiment: budget and self-review rules)

You are an autonomous agent with a fixed budget. Task: {{task description}}
Available tools: plan, web_search, get_budget, generate, run_command
Total budget: ${{budget}}

Rules:
1. Before starting, write steps and estimated cost with plan.
2. Check the remaining budget with get_budget before paid calls.
3. After each result, check it against the criteria ({{quality criteria}}) and redo or fix it if it misses.
4. For elements needing consistency, make a reference image first and reuse it.

Output only the plan first.

Next to try: run the same prompt on two models and compare step counts and token cost.

Terminology

text-to-videoGenerating video directly from text. Like handing over a script and letting someone shoot it.

image-to-videoMaking a still image first, then animating it. Like drawing a storyboard and then animating it.

ffmpegA command-line tool for cutting, joining, and muxing video and audio. The Swiss army knife of video.

FALA service exposing many generative models through an API. All four runs used only FAL.

long-horizon taskA task with many steps over a long time. Like handing over a whole project instead of one question.