Quick answer: 960P outputs 1728x960 at 3 credits per second, the same rate as 1080P, with a fifteen-second ceiling instead of standard 1080P's ten. Multimode 1080P also runs up to fifteen seconds. 960P is image-to-video only and rejects a seed outright.
MiniMax H3 is sold in six output tiers on Callirra. Five of them are easy to reason about: 480P and 768P for drafts, 1080P for delivery, 1440P for preset-driven work, 2K for the wider feature set. Then there is 960P, which sits between 768P and 1080P, costs the same as 1080P, produces fewer pixels than 1080P, and runs for fifteen seconds without selecting a multimode style preset.
That combination is unusual enough that it deserves its own post, because most write-ups skip it entirely.
The numbers
| 960P | 1080P | |
|---|---|---|
| Output | 1728×960 | 1920×1056 |
| Rate | 3 credits / s | 3 credits / s |
| Maximum duration | 15 s | 10 s standard; 15 s multimode |
| Text to video | no | yes |
| Image to video | yes | yes |
| Aspect ratios | 16:9, 9:16 | 16:9, 9:16 |
| Reference images | up to 9 | up to 9 |
| Reference audio | up to 3 | up to 3 |
| First + last frame | no | no |
| Seed | rejected | optional integer, 1 or above |
Two facts carry the whole case: the same three credits per second buys fifty percent more runtime than standard 1080P on 960P, and 960P is image-to-video only. Multimode 1080P is another fifteen-second option.
A third detail people hit late: 960P is the one tier that refuses a seed outright rather than ignoring it. If your pipeline passes a seed to every job, it will be rejected here — and since the lane exists for reference-led work, "no seed" means reproducibility comes from the reference image, not from a number.
Why a "lower" tier costs the same
1728×960 is 1.66 megapixels; 1920×1056 is 2.03. So 960P is not a cheaper raster being sold at the same price — it is a different conditioning pipeline aimed at reference-led work, which is why the product calls it cinematic quality rather than HD.
If you go in expecting "1080P but smaller", you will be confused. If you go in knowing it is the lane for moving a reference still for longer, the pricing makes sense: you are paying for the reference conditioning and the extra duration headroom, not for pixels.
When 960P is the right choice
Your shot needs to run past ten seconds. A fifteen-second continuous take — a slow push through a room, a single unbroken character beat, a loop for a hero section — exceeds standard 1080P's ten-second ceiling. On 960P it is a normal job; multimode 1080P is an alternative if you want its style presets.
You are animating a still with references. 960P is image-to-video only, with up to nine reference images and three audio clips. If your workflow already starts from a frame, you lose nothing by choosing it.
You want a delivery-adjacent tier without paying 2K money. 1728×960 upscales to 1080p cleanly if a downstream system insists on 1920×1080 — you are inventing fewer rows than you would at 768P.
When is 960P the wrong choice?
You need text to video. 960P has no text-to-video scene. A prompt-only job will not run here; use 768P or 1080P.
You are delivering at true 1080p. If the client spec says 1920×1080, account for the measured 1920×1056 standard 1080P output as well as the smaller 960P raster. Check the returned file and resize or pad for the delivery specification.
You need first and last frame. Not supported at this tier, or at 1080P, or at 1440P. Only 480P/768P and the 2K image-to-video lane take both frames.
A practical pairing
Use 768P to find the take, then re-render the selected shot at 960P when you want its reference-led lane, or at 1080P when you want that lane's output. Standard 1080P stops at ten seconds; multimode reaches fifteen. Both tiers cost three credits a second, so choose the lane and check the returned raster against your delivery needs.
FAQ
Is 960P cheaper than 1080P? No. Both are three credits per second. 960P runs longer than standard 1080P; multimode 1080P also reaches fifteen seconds.
Does 960P accept text-to-video? No. It is image-to-video only, so a prompt-only job will not run here.
Why was my seed rejected on 960P? The tier rejects a seed outright rather than ignoring it. Reproducibility comes from the reference image.
Where this runs
All six tiers, with output dimensions and the ten-second cost for each, are on the rate card at https://callirra.com/minimax-h3. The generator exposes 960P as 960P (Cinematic quality) and only offers it in image-to-video mode, which is a deliberate restriction rather than a missing feature.
I work on Callirra, so these numbers are the published rates rather than an independent benchmark. The registry that decides which tier accepts which scene is a single table in our source; if a specific combination rejects your request, that table is the reason, and the error message names the tier and the scene.