I typed a nice prompt into 960P and got an error back.
Not a bad render. A rejection. No text prompts on that tier, at least not on its own: it wants a picture before it will move anything. I had assumed every tier was a size, and this one turned out to be a workflow.
The short version. 960P renders 1728×960 at $0.03 per output second, requires at least one reference image, rejects seeds, and accepts mp3 or wav audio only, up to 15 seconds.
What is 960P?
It renders 1728×960 at three cents per second — exactly the 1080P rate — and it is image-to-video only. You supply at least one reference image. It does not accept a seed. And the audio list is narrower than the standard tiers: mp3 or wav, no flac.
The two differences that matter in practice are the run time and the missing seed. It gives you fifteen seconds where 1080P stops at ten, and it takes away the ability to hold the random start constant while you change one word.
Why the reference is not an obstacle
I stopped resenting it around the third take. The reference supplies identity — the face, the label, the fabric, the light on it — and the tier supplies the motion and the surface. Ask it for something from nothing and you get a render with nothing to anchor it, which is precisely the look people dislike in AI footage.
My whole loop changed after that: make a still I already like, at whatever image tier is cheapest, and bring it in. The prompt becomes a note to the animator instead of a casting call.
Why can a seed be used everywhere except here?
Seeds are how I compare. Hold the seed, change one word; hold the words, change the seed. Without that, every comparison moves two variables and I learn nothing.
So I do the exploration somewhere else. Draft on a tier that accepts a seed, find the brief that works, then finish the approved take on 960P. It is one extra step and it prevents the thing I actually hate, which is not knowing whether a bad take was the prompt or the roll.
The audio detail that costs a round trip
mp3 and wav only. If the voice reference came out of a DAW as flac, convert it before you drag it in. The job is rejected before it renders, so you lose a click instead of credits — but you do lose the walk to the kitchen and back.
Audio references are still free here, up to three clips of two to fifteen seconds.
When is 960P the wrong tier?
If you are exploring from words, this is an expensive way to think. If you need a square canvas, it does not exist here. If your comparison method depends on seeds, it is not the tier to compare on. And if the shot needs more than fifteen seconds, no tier will help you — plan two takes.
The honest version: 960P is not "better 1080P". It is a reference-led path for shots that must look photographed, and the extra constraints are part of what buys that look.
The measured 960P takes and the full rate card are on the H3 page.
I work on Callirra, which is why I can see the rejection rules rather than guess them, and why my sample of takes lives in one place. If you have reference-led output from another H3 lane that behaves differently, the interesting part is the reference you used, not the model name.