The pick is a method, not a leaderboard.
Model rankings rot in weeks. Dated Sep 2026 picks for UGC-style shorts, plus the per-brief selection method that outlives any leaderboard.
In the Sep 2026 snapshot below, Veo 3.1 is the pick when the brief says commercial, Sora when it says story, and an image-route hybrid when the same character has to carry many ads.
Route your brief
Three questions from the selection method below, applied for you.
Directional routing, not a leaderboard. Dated picks live in the table below.
Dated, not permanent
Strengths per use case, from our own runs and documented public work.
| Model | Wins at | Observed limits | Pick it when |
|---|---|---|---|
| Veo 3.1 | Cinematic ad looks, image-to-video consistency, native audio that usually lands on the first route | Prompt discipline required for identity hold across cuts; cost climbs with retries | The brief says "commercial". Polish, brand-safe framing, cinematic camera moves. |
| Sora | Generative storytelling, surreal and stylized hooks, strong scene coherence on longer shots | Realism can drift into uncanny for testimonial-style UGC; stylized output needs disclosure discipline | The brief says "story". Scroll-stopping concepts where polish matters less than the idea. |
| Image-route hybrids | Personalized ads anchored on a reference image; character consistency via stills-then-animate | Two-step pipeline adds latency and its own failure modes | The brief says "same character, many ads". Identity consistency beats everything else. |
Four questions, in order
They settle the choice faster than any leaderboard.
1. Does the character need to persist?
Same face across six exports? Take the image-route hybrid: lock the character in stills first, animate second. Free-form models will not hold identity across a 6-export set reliably.
2. Is it an ad or a story?
Ads need commercial framing and clean audio: Veo 3.1's home turf. Stories need an idea that survives low fidelity: Sora's. Picking the wrong lane is the most expensive mistake in the pipeline.
3. What does a retry cost?
Our pilot ran 47 metered calls with 3 concepts surviving to export. Every failed generation is spend. Models with weak first-route hit rates cost more than their list price.
4. What did the last run's QA say?
Our release gate scores every export 0 to 100 on decode, loudness, black-frame, freeze, fast-start. The 16/16 campaign pass and the 91 to 96 pilot scores are dated receipts for the routes we picked. Your QA log beats any influencer's ranking.
Three fronts where models leapfrog
Version numbers move. Audio sync quality, identity hold, and price per successful export are the three fronts where the models leapfrog each other right now. The picks above are checked each week in the club before any brief routes to them, which is why the picks inside the club carry dates and the reasoning, not just names.
The camera is optional. Taste is not.
One brief, one finished roll, critique each week. Dated picks, checked weekly in the club. Sources checked 24 Sep 2026.