Ten Video Models, One 60-Second Brief: A 2026 Stress Test With Receipts
At Hedra AI, we like generative media research to leave marks. So for this test we wrote one 60-second short-film brief, froze the prompt pack, set a hard budget, and ran it through ten video models that US creators can reasonably get their hands on in 2026. No cherry-picked hero clips. No manual cleanup that hides the model's weaknesses. Just the same brief, the same shot list, and a spreadsheet that got mean.
The brief: Last Bus on Route 9
Logline: A night-shift nurse in Chicago misses the last bus and finds a humming package on the bench. One location. Two characters. One practical light source. One line of dialogue. A short chase beat. A quiet ending. Runtime: 60 seconds. Aspect ratio: 16:9. Style: neo-noir, sodium-vapor streetlights, wet pavement, no sci-fi gloss.
We broke it into 12 shots of five seconds each. Each model got the same shot descriptions, the same character reference image for image-to-video runs, and up to three retries per shot. We capped spend at $250 per model and assembled everything in DaVinci Resolve with cuts, fades, and a licensed music bed only. No generative fixes in post.
The prompt skeleton
Shot [N/12]: [action]. Character A: [fixed description]. Character B: [fixed description]. Location: [fixed description]. Camera: [lens, height, movement]. Lighting: [practical source, contrast]. Audio: [dialogue line or ambience]. Duration: 5s. Style: [neo-noir](https://en.wikipedia.org/wiki/Neo-noir), realistic, no text, no logos.
We also gave every model the same negative prompt: extra fingers, warped face, floating limbs, text, watermark, jump cuts, inconsistent wardrobe. For image-to-video models, we used the same two character stills and one location plate.
How we scored
Six axes, each 1-5, weighted for a small production team:
- Character consistency: face, wardrobe, body proportions across shots.
- Camera control: did the model respect lens, height, and move?
- Audio sync: dialogue and Foley alignment when native audio was available.
- Latency: median time from request to usable clip.
- Cost: dollars per usable second after retries.
- Licensing clarity: whether commercial use, training-data provenance, and output ownership were easy to confirm.
We also tracked a usable-shot rate: the percentage of generated clips we could put in a timeline without regenerating.
The scoreboard
| Model class | Consistency | Camera | Audio sync | Latency | Cost | Licensing clarity | Usable-shot rate |
|---|---|---|---|---|---|---|---|
| Hosted flagship A (text-to-video) | 3.0 | 4.5 | 4.0 | 3.0 | 2.0 | 3.0 | 58% |
| Hosted flagship B (image-to-video) | 4.5 | 4.0 | 3.0 | 3.5 | 2.5 | 3.0 | 72% |
| Native-audio model C | 3.5 | 3.5 | 5.0 | 2.0 | 1.5 | 2.5 | 61% |
| Open-weight model D | 2.5 | 3.0 | 2.0 | 4.5 | 5.0 | 4.5 | 49% |
| Open-weight model E | 2.0 | 2.5 | 1.5 | 4.0 | 5.0 | 4.0 | 41% |
| Prosumer model F | 3.5 | 4.0 | 2.5 | 4.0 | 3.5 | 3.0 | 66% |
| Prosumer model G | 3.0 | 3.0 | 3.0 | 4.5 | 4.0 | 3.5 | 63% |
| Mobile-first model H | 2.0 | 2.0 | 2.0 | 5.0 | 4.5 | 2.0 | 34% |
| Enterprise model I | 4.0 | 4.5 | 3.5 | 2.5 | 1.0 | 4.0 | 69% |
| Research preview J | 2.5 | 3.5 | 3.0 | 3.0 | 3.0 | 2.0 | 52% |
Scores are averages from three reviewers. Full clip-by-clip notes and the named model map are in the Hedra AI lab notebook.
What actually broke
Character consistency is still a shot-count problem. Every model held the lead character for the first two or three shots. By shot eight, jackets changed color, jawlines drifted, and one model gave the nurse a different hairstyle in the same scene. Image-to-video beat text-to-video here, but only when we supplied clean reference stills and kept camera moves simple.
Camera control is better than it was, but it is not a director's monitor. Models followed slow push in and handheld follow more reliably than 35mm, chest height, slight Dutch angle. If a shot depended on precise blocking, we burned retries.
Native audio is a trap and a gift. The models with audio sync saved us a Foley pass, but they also invented dialogue we did not ask for and drifted lip sync on close-ups. If you need a specific line reading, generate picture first and treat audio as a separate department.
Cost per usable second is the number that matters. The cheapest raw generation was not the cheapest workflow. Open-weight models won on sticker price and licensing clarity, but they ate time in retries and post. Hosted flagships cost more per clip and less per usable shot in some cases.
Licensing clarity is still a PDF scavenger hunt. A few vendors made commercial use obvious. Others buried training-data provenance in a help center article. For US creators, that is not a footnote. It is a distribution risk.
The decision guide
- If you need a dialogue-driven 60-second piece with tight turnaround: pick a hosted image-to-video model with native audio, budget for 2.5x retries, and lock character references before generating.
- If you need maximum camera control: use a hosted flagship with camera-move controls and keep shots under five seconds.
- If you need low cost and clear licensing: use an open-weight model, but plan for a human editor and a color pass.
- If you need a pitch or lookbook: mobile-first and research previews are fine. Do not promise final delivery.
- If you need US broadcast or brand work: verify output ownership, training-data terms, and indemnification before the first prompt.
Reproducibility notes
We ran the test in October 2026 on US-based accounts. We paid for every credit. We did not use vendor-sponsored trials. We published the prompt pack, the shot list, the scoring rubric, and the raw reviewer notes. We anonymized model names in the public scoreboard because several vendors' terms restrict publishing raw outputs, but the internal ledger is available to verified researchers.
The takeaway
Ten models can make a beautiful eight-second clip. Fewer can make a coherent 60-second film. The gap is not magic. It is consistency, control, and paperwork. Pick the workflow that matches your weakest link, not the demo that made you gasp.
