July 16, 2026 — TryAI published the Music Video Arena: give Claude Fable 5 and GPT-5.6 Sol the same song, the same tool belt, and a $25 or $100 generation budget — then walk away. The agents planned shots, searched the web, burned FAL credits on images and video, and stitched outputs with ffmpeg. The Hacker News thread (~146 points) and the open github.com/hershalb/music-video-arena repo make this one of the clearest autonomous creative agent experiments of the summer — and the verdict is uniformly interesting, not good.
TL;DR — four runs, one song
| Run | Wall time | Gen artifacts | FAL/gen spend | Resolution | Total cost (incl. LLM) |
|---|---|---|---|---|---|
| Fable $25 | 39 min | 54 videos | $24.30 | 1280×720 | ~$30+ (see blog) |
| Fable $100 | Longer | More clips | Higher FAL | Mixed | $73.65 (most expensive) |
| Sol $25 | 42 min | 61 images + 46 videos | Budget-capped | Mixed | $27.45 (cheapest total) |
| Sol $100 | Longer | Mixed Wan/Veo/Hailuo | Higher FAL | Mixed | Mid-tier total |
Song for all runs: Bruno Mars / Mark Ronson — "Uptown Funk" (same audio track, same lyrics challenge).
| Question | Answer |
|---|---|
| Human in the loop? | No mid-run — autonomous harness only |
| Tools exposed? | plan, web_search, get_budget, generate_image, generate_video, run_command |
| Video backend? | FAL APIs (Wan, Seedance, Veo, Hailuo — model-dependent) |
| Quality verdict? | Awful but informative — HN + TryAI agree |
| Open source? | Yes — hershalb/music-video-arena |
The harness — what "autonomous" means here
TryAI's setup is a textbook agent harness:
Song + budget cap
│
▼
┌──────────────────┐
│ Frontier LLM │ Fable 5 or GPT-5.6 Sol
│ (planner) │
└────────┬─────────┘
│ tool calls
┌────┴────┬────────────┬─────────────┐
▼ ▼ ▼ ▼
plan web_search generate_* run_command
│ │
FAL image/ ffmpeg stitch
video
| Tool | Role |
|---|---|
plan | Shot list, lyric mapping, budget allocation |
web_search | Reference gathering (costumes, era, artist visuals) |
get_budget | Remaining dollars — forces trade-offs mid-run |
generate_image | Keyframes / character stills |
generate_video | Clip generation via FAL |
run_command | ffmpeg concat, trim, basic post |
No human corrected a bad take. No director re-cut on beat drops. That is the point — and the failure mode.
Results — what went wrong (predictably)
TryAI and HN commenters catalogued the same quality ceiling:
Literal lyrics beat rhythm
Agents interpreted lyrics literally instead of musically:
- "Michelle Pfeiffer" → random blonde woman in gold
- "Uptown" → generic city b-roll without funk choreography
- Tempo sync — cuts rarely landed on downbeats; ffmpeg concat ≠ editing
This mirrors long-form continuity problems in ViMax-style agentic video — short clips hide sync issues; a full song exposes them.
Character consistency failed
No run maintained a stable lead performer across scenes. Image-to-video pipelines drift faces, wardrobe, and lighting — the same class of bug Seedance / Kling production posts warn about for 30-second clips, compounded across 3–4 minutes of music.
No self-review loop
Neither Fable nor Sol reliably rejected bad clips before spending the next dollar. A human producer would bin 80% of takes; the harness treated generation success as usable footage.
Budget ≠ quality
| Observation | Implication |
|---|---|
| Fable $100 = $73.65 total | Most expensive — not best MV |
| Sol $25 = $27.45 total | Cheapest — still not good |
| Fable $25 = 54 videos in 39m | High clip velocity, low curation |
For token economics context, compare GPT-5.6 vs Fable 5 on coding benchmarks — creative spend does not inherit coding leaderboard ordering.
Video model choices — Wan, Seedance, Veo, Hailuo
Agents picked models inside the tool API — not a fixed pipeline.
| Budget | Fable 5 (reported) | GPT-5.6 Sol (reported) |
|---|---|---|
| $25 | Wan 2.5, Seedance via FAL | Image-to-video pipeline |
| $100 | Mixed Wan, Seedance, other FAL endpoints | Mixed Wan, Veo, Hailuo |
Seedance context: ByteDance's model family appears in explainx.ai's Seedance 2.5 guide and Odyssey film coverage — strong for short 4K clips, not automatic music-video grammar.
Broader tool survey: AI video generation complete guide.
Fable vs Sol — agent behavior, not MV beauty
| Dimension | Fable 5 | GPT-5.6 Sol |
|---|---|---|
| $25 run time | 39 min | 42 min |
| $25 artifact count | 54 videos | 61 images + 46 videos |
| $100 total cost | Highest ($73.65) | Mid |
| $25 total cost | Higher than Sol | Lowest ($27.45) |
| Coding benchmark prior | SWE-Bench Pro leader | Terminal-Bench / ALE strengths |
| This arena | More FAL burn at $100 | Cheaper at $25, same bad MV |
Takeaway: Frontier coding rankings do not transfer to autonomous MV direction. The harness and lack of human taste loop dominate.
HN consensus — awful but interesting
Hacker News treated Music Video Arena as a capability probe, not a product launch:
- Autonomous agents can spend money — budgets, FAL, ffmpeg — end to end
- Creative quality flatlines without human taste or beat-aware editing
- Open sourcing the harness matters more than any single video output
- Human-in-the-loop still wins for anything you'd show an audience
Parallel debate: AI copying creativity — autonomous MVs add volume without original direction, amplifying slop risk unless C2PA / provenance labels ship on outputs.
Reproduce or extend the benchmark
Clone the official repo and swap models:
git clone https://github.com/hershalb/music-video-arena.git
cd music-video-arena
# Follow README — API keys for FAL + frontier LLM providers
Ideas for reruns explainx.ai would track:
- Different song structures (ballad vs EDM vs rap) — tempo sync stress
- Human checkpoint every N clips — measure quality delta vs cost
- Open-weight planner (Kimi K3 API) vs Fable/Sol — separates harness from model
- Fixed video model — isolate Wan 2.5 vs Seedance vs Veo
Pair with loop engineering if you add verification subagents ("reject clip if face drift > threshold").
What production teams should actually do
| Arena lesson | Production implication |
|---|---|
| Literal lyric visuals | Storyboard before agent runs; lock metaphor bible |
| No beat sync | Human editor or DAW-marked timestamps in harness |
| Character drift | Locked reference stills + img2vid only from approved frames |
| No clip QA | Add self-critique tool or second-pass reviewer model |
| Budget runaway | Cap per-shot spend, not just total — Fable burned fast at $25 |
For agentic long-form ambition, read OpenMontage and Meta Muse video agents — same continuity wall, higher production gloss.
Summary
TryAI Music Video Arena (July 16, 2026) ran Fable 5 and GPT-5.6 Sol at $25 and $100 on "Uptown Funk" with plan → FAL → ffmpeg autonomy. None were great — literal lyrics, bad sync, inconsistent characters, no self-review. Fable $100 cost the most ($73.65 total); Sol $25 cost the least ($27.45). The value is the open harness and the proof that human-in-the-loop still wins for watchable music video. Use gpt-5-6 vs Fable for coding procurement; use this arena for creative agent economics.
Related on explainx.ai
- TryAI's Canvas Arena: GPT-5.6 Sol vs. Fable 5 vs. Grok 4.5 drawing
- GPT-5.6 vs Claude Fable 5 comparison
- Fable 5 open-source enterprise alternatives
- What is an agent harness?
- AI video generation — Sora, Runway, Kling guide
- Seedance 2.5 — 30-second 4K AI video
- ViMax agentic video generation
- AI copying creativity — shadcn debate
- LinkedIn C2PA content credentials
- Loop engineering for coding agents
- GPT-5.6 Sol, Terra, Luna preview
Official sources: TryAI Music Video Arena blog · github.com/hershalb/music-video-arena
Run timings, costs, and model names accurate as of July 17, 2026 per TryAI blog and HN discussion. FAL pricing and model IDs change — verify README in the repo before rerunning.