TryAI gave two frontier models $100 each and six tools — plan, web search, budget tracking, image generation, video generation, and ffmpeg — and told them to make a music video for "Uptown Funk." The resulting videos don't look good. That's not the story. What's interesting is how each model spent its money, and what that reveals about the autonomous decision-making baked into their design.
Claude Fable 5 consistently burned 30–40% of its budget on LLM inference alone — $16.99 on the $25 run, $25.05 on the $100 run. GPT-5.6 Sol spent about 10% on tokens, capping out at $4.27. That's a 3–6x gap in model overhead on the same task. The difference isn't intelligence — it's architecture.
| Run | Total cost | LLM cost | LLM share |
|---|---|---|---|
| Fable 5, $100 budget | $73.65 | $25.05 | 34% |
| Sol, $100 budget | $39.82 | $3.25 | 8% |
| Fable 5, $25 budget | $41.29 | $16.99 | 41% |
| Sol, $25 budget | $27.45 | $4.27 | 16% |
The mechanism: GPT-5.6 Sol cached roughly 87% of its input tokens across the run — repeated tool call schemas and shared context don't get re-billed. Claude Fable 5 has no equivalent prompt caching in this setup, so every turn pays full price for everything. That alone reverses the usual in-house cost equation. For any team running agents at scale, token reuse architecture is now a first-order economic variable, not an optimization footnote.
Reliability vs. Exploration Is a Cost Decision
The models diverged on strategy, not just price. Claude Fable 5 committed to a single text-to-video model per run and never changed: Wan 2.5 at the $25 tier, Seedance 1.0 Pro at $100. Zero failed generation calls on the high-budget run. Clean, predictable, pipeline-shaped.
GPT-5.6 Sol at $25 built an image-to-video pipeline — FLUX schnell for stills, then Wan 2.2-5b for animation — overlaying text and applying video effects. It was the most inventive run in the experiment. It also had 10 failed calls (transient provider errors it retried through). At $100, Sol switched between three different video models (Wan 2.5, Veo 3.1 Lite, Hailuo 2.3) and produced lower quality clips than its cheaper run. More options, worse outcome.
This maps to a pattern I've seen with agent orchestration across other tools (recent work on PlanWright covered similar orchestration trade-offs). A model that defaults to reliability will converge on one approach and execute cleanly. A model that defaults to exploration will retry, combine, switch — and sometimes find a better result, sometimes burn runway. The cost structure of the underlying LLM (cheap tokens on Sol, expensive tokens on Claude) directly enables or constrains which strategy is viable. Sol can afford to experiment because its cognitive overhead is low. Claude Fable 5 can't, so it doesn't try.
The Missing Loop: Nobody Reviewed Their Own Work
Here's the most revealing failure, and it's not about cost. Across all four runs, neither model reviewed a single frame. No self-critique, no re-cut, no second pass. Characters didn't stay consistent. The lyrics were taken literally (credits sequence for "Uptown Funk"? Nobody generated a dance move). Tempo matching was nonexistent. The videos are a concatenation of first-pass clips, assembled once and shipped.
These are frontier models with multi-million-token context windows and tool-use loops. The harness gave them run_command with ffmpeg — they could have split clips, re-rendered, overlayed, anything. They just … didn't. The autonomy is real but shallow: the models solve the planning problem, but they don't evaluate the output problem. We're missing a metacognitive step — "look at what you just built and decide if it's worth keeping" — and no pricing tier makes a model add that step on its own.
The open-source harness (github.com/hershalb/music-video-arena) is worth running yourself if you want to see the full logs. The data are clean, the comparisons are fair, and the verdict is useful: if you're building agent workflows, design for the model you're using. Claude Fable 5 needs explicit guardrails against token waste — maybe a max-step limiter or a caching layer GPT already has. GPT-5.6 Sol needs harder quality gates, because it will happily generate 10 bad clips at $0.40 each and call it done.
The next 10x improvement in agent outputs won't come from a better model. It will come from an agent that watches its own output and decides to try again.