Claude Fable 5 vs GPT-5.6 Sol: $100 AI music video test

Key Takeaways

• Claude Fable 5 spent $48.60 on generation but $25 on tokens, totaling $73.65. GPT-5.6 Sol spent $36.57 on generation but only $3.25 on tokens, totaling $39.82.
• Both models autonomously researched video APIs, generated clips, edited with ffmpeg, and muxed final videos without human intervention.
• GPT-5.6 Sol experimented with multiple video models in one run while Claude stuck to a single approach per budget tier.

A developer experiment pitted Claude Fable 5 against GPT-5.6 Sol in an unusual test: hand each model a song, a fixed dollar budget, and a set of tools, then let it autonomously produce a complete music video. No human editing. No hand-holding. The models researched video generation APIs, generated clips, analyzed their own footage, ran ffmpeg commands, and assembled final cuts. The result is a rare look at how frontier AI models make creative decisions when given real constraints.

What the experiment actually tested

The team at TryAI.dev built an agentic harness with six tools available to each model. A planning tool for thinking through strategy (free). Web search to research which video generation models exist. Budget checking. Image and video generation via FAL or Replicate APIs. And a local shell with ffmpeg for editing.

Both models received identical inputs: Bruno Mars and Mark Ronson's "Uptown Funk," a short text description, and a time-stamped lyric transcript. Each model ran at two budget levels, $25 and $100, for four total runs. Once the budget hit zero, paid generation stopped, but the model could continue editing whatever footage it had already produced.

The cost breakdown tells a different story than the headline

The $100 budget caps only the generation spend, meaning what the model pays FAL for actual video creation. But running the model itself costs money too. Claude Fable 5 charges $10 per million input tokens and $50 per million output tokens. GPT-5.6 Sol charges $5 and $30 respectively.

Claude produced more clips at the $100 tier (80 vs. 70) and output at 1920x1080 resolution. But it burned through nearly eight times more on token costs than GPT-5.6 Sol. The Sol model benefited heavily from cached inputs, with 1.8 million of its 2.1 million input tokens coming from cache at the $100 level.

Different models made different creative choices

Given freedom to pick their own tools, the models diverged. Three of four runs went pure text-to-video. Only GPT-5.6 Sol at the $25 budget used an image-to-video pipeline, generating still frames with FLUX schnell at $0.003 per image, then animating them with Wan 2.2-5b.

At $100, GPT-5.6 Sol mixed three different video models in a single run: Wan 2.5 at $0.05 per second, Veo 3.1 Lite at $0.10 per second, and Hailuo 2.3 Standard at $0.28 per video. Claude Fable 5 stuck with a single model at each budget tier. At $100, it used Seedance 1.0 Pro, which outputs 1080p video at roughly $0.12 per second.

The error rates varied too. Claude Fable 5 at $100 had zero failed generation calls. GPT-5.6 Sol at $25 had 10 failures, mostly transient network errors. Failed calls weren't charged, but they consumed steps and time.

Wall-clock time was surprisingly similar

Despite different strategies, all four runs finished in 38 to 50 minutes. Claude Fable 5 at $100 finished fastest at 38 minutes 56 seconds with 28 steps. GPT-5.6 Sol at $100 took longest at 49 minutes 39 seconds with 34 steps. The $25 runs landed in between.

All four runs completed autonomously. None hit a step limit or timeout. Each produced a valid full-length video with the original song muxed in.

What founders should take from this

The experiment reveals a gap between nominal capability and actual cost. A "$100 AI video" might cost $73 or $40 depending on which model you choose, and the difference comes almost entirely from token pricing, not creative output. For startups building agentic workflows, model selection based on per-token cost matters more than headline budgets.

The models also showed different risk tolerances. Claude stuck with proven approaches. GPT-5.6 Sol experimented more, mixing models and trying image-to-video pipelines. Neither is inherently better. The right choice depends on whether you're optimizing for predictability or exploration.

The full harness is open-sourced at github.com/hershalb/music-video-arena. Anyone can fork it, swap in different models, and run their own comparisons.

Frequently Asked Questions

How much does it cost to generate an AI music video?

In this test, total costs ranged from $27.45 to $73.65 for a full-length video, depending on the model and budget tier. Generation APIs charged $0.05 to $0.28 per clip, but the model's own token usage added $3 to $25 per run.

Can AI autonomously edit video with ffmpeg?

Yes. Both Claude Fable 5 and GPT-5.6 Sol used ffmpeg commands to analyze audio, cut clips, concatenate footage, and mux the final video with the original song. The models wrote and executed these commands without human intervention.

Which AI model is better for video generation?

It depends on your priorities. Claude Fable 5 produced more clips and higher resolution output but cost nearly twice as much overall. GPT-5.6 Sol experimented with multiple video models and finished cheaper due to lower token costs.

What video generation APIs did the models use?

The models chose from FAL and Replicate APIs. Options included Wan 2.5, Wan 2.2-5b, Veo 3.1 Lite, Hailuo 2.3 Standard, Seedance 1.0 Pro, and FLUX schnell for image generation.

Source: Hacker News: Best / TryAI