Higgsfield, tested: what the credits actually buy

Higgsfield does one thing better than anything else we have tested: it moves the camera. The orbit that looks like it was shot on a crane, the freeze-and-rotate, the push-in that holds — those come out right more often here than anywhere else, and they come out right because you name the move instead of describing it.

It is also the platform where it is hardest to know what you are spending. This page is what we found after running nine generations and reading its parameters off the API, on 22 October 2026.

What it actually is

Two things at once, and almost every review treats it as one.

It has its own models — Cinema Studio, in three versions, plus Marketing Studio for product ads. And it resells other companies' models through the same interface: FLUX 3 from Black Forest Labs, Grok 1.5 from xAI, Hailuo and H3 from MiniMax, three versions of Seedance from Bytedance, Wan 2.6.

That matters because the parameters, the maximum duration and the way you prompt all differ between them. "How to prompt on Higgsfield" is not a thing; "how to prompt Cinema Studio v2" is.

What a generation costs

These are measured, not estimated — we asked the API to price each one before submitting it.

Generation	Credits
Cinema Studio v2, 3 seconds, standard mode	3
FLUX 3 Video, 5 seconds, 720p, with audio	30
FLUX 3 Video Edit	1 per second of the processed clip

The ten-to-one gap between the house model and FLUX 3 is the single most useful number on this page. If you are iterating, iterate on Cinema Studio.

Three things move the price: resolution (the docs say plainly "higher = more credits"), duration, and the model. And the default resolution is 720p, which is already not the cheapest setting — if you have never touched it, you are generating above the floor.

Two things the documentation promises that did not work

This is the part you will not find in other reviews, because finding it costs credits.

Multi-shot does not let you choose the shots. Cinema Studio v2 has a multi_shots switch that splits one generation into several shots. We turned it on, wrote a three-shot list into the prompt, and got back an unrelated science-fiction scene — the prompt was discarded entirely. With the switch off, the same prompt was followed exactly.

The model's own documentation says the shots are "driven by multi_prompt", so we passed that. The platform refused it: "Cinema Studio Video does not support this parameter." The field that tells the model what the shots are is not available on this access path, so the switch puts the model in multi-shot mode with no shot list, and it invents a scene.

Audio defaults are inconsistent between models. Cinema Studio 3.0 generates silence by default; v2, FLUX 3 and the Seedance family generate audio by default. If you move between models, audio behaviour changes under you — and if you score your videos in an edit, you are paying to generate audio you throw away, on every test take.

What worked better than expected

Dubbing. One call, a target language, and it comes back translated, voiced and lip-synced. We dubbed a five-second English clip into Portuguese and checked it three ways: machine transcription of both audio tracks confirmed the translation was exact, the duration was identical, and the mouth shapes differ between the two videos at the same frame numbers — the lip sync genuinely ran. Eighteen languages are supported.

One caveat that belongs in your schedule rather than your expectations: it is slow. Ours sat in a queue for several minutes before processing even began. Eighteen languages is eighteen queued jobs, not eighteen instant conversions.

Named effects. Dozens of fixed motion presets — BULLET TIME, EARTH ZOOM, TRACKING, FROZEN IN MOTION. Picking one beats describing the same move in prose, and the platform agrees: when we submitted a prose prompt describing an orbit, it declined to generate and suggested a preset instead.

Its own course is free, and so is the certificate. Eight lessons, 1h42m, beginner level, and the certificate costs nothing — which is rarer than it sounds. Most free courses sell the paper.

The free plan, and an honest gap

Our catalogue has recorded the free tier as 10 credits a day with a watermark, checked on 15 August 2026. We could not re-verify that on 22 October: the pricing page renders behind a login, and we are not going to republish a two-month-old number as if it were current.

So treat the free tier as "exists, limited, watermarked" and check the figure yourself before you plan around it. When we can confirm it, this line changes.

Who it is for

Worth paying for if camera movement is the point of your work — product shots that need to feel shot, openers, anything where the move carries the idea. Nothing else we have tested directs motion this reliably.

Not worth it if you mainly need a talking head, a long clip or many shots in sequence. The house models stop at 15 seconds, multi-shot is a lottery right now, and a face in close-up is better served elsewhere.

And budget the practice, not just the subscription. The teaching is free; the generating is not. That is the honest shape of the cost.