I counted six "I tested every AI video model so you don't have to" articles published last week alone. Same format. Same models. Same conclusion: it depends. They run the same prompt through Kling, Runway, Veo, Sora, and Seedance, line up the outputs, pick a winner, and collect their clicks.
The articles are fine. The question is broken.
The camera store problem
"Which model is best" is the new "which camera should I buy." It haunted photography forums for twenty years. People would spend months comparing sensor specs and dynamic range charts instead of going outside and taking a photograph. The answer was always the same: the best camera is the one that matches how you see. Nobody wanted to hear it then either.
AI video models are not interchangeable commodities with different price tags. They are different instruments. They see differently. They move differently. They have preferences they did not choose and habits they cannot break. Ranking them first to fifth on a single prompt is like ranking five DPs on a single setup. You will get a result. It will tell you almost nothing useful.
Models have temperaments
Eleven articles in this series and a pattern keeps surfacing. Every model has a personality that persists across every category we have tested. Movement, color, lighting, sound, structure. The personality is the same.
One model is a control freak. Give it twelve constraints and it will attempt all twelve. It will not add what you did not ask for. It will not improve your prompt. It is a precision instrument, and if you do not like the result, that is on you.
Another is an aesthete. It has opinions about what looks good and it will gently overrule you to get there. Hand it a deliberately ugly lighting setup and watch it smooth the shadows toward something it finds more appealing. This is infuriating if you wanted ugly. It is a gift if you want beauty without specifying every parameter.
A third is a storyteller. It does not care about pixel-level fidelity to your reference frame. It cares about whether the shot feels like part of a narrative. Give it a sequence and it will drift visually in ways that serve the story even when those drifts violate your specifications.
A fourth is a preservationist. It holds what you give it. Color does not drift. Faces do not shift. Wardrobe does not change. If you need consistency above all else, it will give you consistency. If you wanted it to riff, you are out of luck.
A fifth is an interpreter. It reads your prompt the way a good actor reads a script, finding subtext, making choices you did not explicitly direct but that feel correct in retrospect. Sometimes those choices are brilliant. Sometimes they are baffling. You will never be bored.
These are not bugs. They are aesthetic signatures. And if you have been reading this series, you already know which model is which. The point is that "best" depends entirely on what you need the shot to do.
The real question is casting
When a director hires a DP, they are not picking the person with the highest score on some imaginary cinematography test. They are picking a sensibility. Roger Deakins sees space differently than Emmanuel Lubezki. Bradford Young lights skin differently than Robert Richardson. You hire the one whose instincts match your vision for this specific project.
Model selection works the same way. The comparison articles treat it as a consumer electronics decision. Megapixels, frame rates, resolution, price. That is the wrong frame. The useful question is not "which model produces the best output" but "which model's instincts match what I need from this shot right now."
Need the model to follow orders? Use the one that follows orders. Need it to have good taste when you do not feel like specifying every shadow? Use the one with taste. Need it to hold a character's face across eight cuts? Use the one that holds faces. Need it to surprise you? Use the one that surprises.
This is casting, not shopping.
Why this matters for prompting
When you understand that models have temperaments, prompting strategy changes. You stop writing one prompt and feeding it to five models to see who wins. You start writing different prompts for different models because you know what each one needs to hear.
The obedient model wants specificity. Give it vague instructions and it will give you a vague result. It has no fallback aesthetic to cover for your laziness. The opinionated model wants headroom. Over-specify and you will fight it. Under-specify and its defaults are gorgeous. The storyteller wants narrative context. "A woman walks" means less to it than "a woman leaves" because the second implies intent, and intent gives it something to interpret.
One prompt does not fit five models any more than one lighting setup fits five scenes. The "I tested the same prompt" methodology is a parlor trick. It tells you which model handled that particular prompt best, not which model is better. Different prompt, different winner. Every time.
The CinePrompt thesis, stated plainly
This is why CinePrompt has a model selector. It is not a dropdown for branding purposes. The prompt it generates changes based on which model you are targeting. Word order shifts. Specificity levels adjust. Certain parameters get emphasized or suppressed. Because the same creative intent requires different language depending on who is listening.
That is not a feature. That is the whole thesis of this tool. A century of cinematography vocabulary is only useful if you are speaking a language the listener understands. The listener varies. So the translation varies. The intent stays the same.
Stop ranking models. Start casting them.
Bruce Belafonte is an AI filmmaker at Light Owl. He has used all five models this week for the same project and does not find this unusual.