Higgsfield is a San Francisco company of Kazakhstani origin, founded in 2023 by Alex Mashrabov, who led generative artificial intelligence at Snap, and by researcher Yerzat Dulat. In October 2025 it became Kazakhstan's first unicorn — that is, the first company in the country valued at over one billion dollars; by August the stated valuation was $5.4 billion.
The product is an aggregator with its own models. It brings together in a single interface generators from other companies — Kling and Seedance for video and Nano Banana for images — and adds its own: Soul 2.0 for images and Soul ID to keep the same face consistent from one generation to the next. On top of the models it layers directorial controls, including Cinema Studio.
Soul ID learns an identity from a few photographs and reapplies it within the Soul 2.0 image model. The company acknowledges variations with unusual angles or significant style changes, and worse results when the source photographs are poor quality.
Every generation attempt has a cost. According to the pricing consulted on September 14, five seconds with Kling 3.0 requires around 7 credits at 720p and 30 at 4K. At the same model and duration, resolution alone more than quadruples consumption. The figures describe a production reality: to estimate the cost of a scene, you have to factor in the discarded attempts too.
The plan's fine print limits "unlimited" use and free generations to the website: it excludes them from MCP and CLI, the access points for agents and commands, and from the Canvas and Supercomputer environments. It also provides for slowdowns during periods of heavy traffic.
The general refund clause allows seven days from first purchase, no credits used, and a fee of up to 6 percent where permitted by law; renewals are excluded. Statutory rights remain in place, as does the specific framework for European and British consumers. At the Better Business Bureau, the US body that collects complaints about companies, 27 have come in over twelve months: very few, roughly one per million declared users. Sixteen went unanswered, and the Better Business Bureau notes that the nature of the complaints and the company's responses matter more than their number. The F grade comes from there.
Higgsfield Sucks, a site dedicated to criticism of the platform, collects testimonials about slowdowns on "unlimited" plans, blocked accounts, and difficulties with refunds and creator payments, and accuses the company of deceptive marketing and improper use of others' work. The accusations are those of its authors.
The Forbes investigation by Rashi Shrivastava, published on February 11, documents promotional kits containing racist videos and archival clips presented as AI-generated. Mahi de Silva, co-founder and chief strategy officer, acknowledged to Forbes that errors had been made in distributing those materials. On the slowdowns and outages, he cited high traffic and fraudulent activity, stating that affected users had been refunded.
To understand the reviews, it helps to look at who gets paid. Higgsfield runs two separate programmes: Earn compensates content produced for campaigns, with a maximum of $2,500 per video; the affiliate programme pays up to 25 per cent on qualifying subscription purchases, for a maximum of twelve months. The Earn ceiling shows how much you can receive, without indicating how many actually reach it.
On Trustpilot, as of 14 September, Higgsfield has 4 out of 5 across 4,351 reviews, with 19 per cent at one star. The listing notes that the company invites customers to leave reviews; competitors that invite no one, such as Runway and Pika, sit between 1.2 and 1.7. Higgsfield has also published, on its own subdomain, a page titled «The Higgsfield billing controversy, explained»: anyone searching for problems with the service finds the company's own response first.
Here at KANSEI Studio we ran into continuity issues using Higgsfield for an interview with Charlie, a generated character. Working with ten-second clips, we got different background tones and intonations between scenes. In one version of the closing shot the framing tightened and the set shifted slightly: we had to replace the take to match it with the previous ones.
The voice came across as too metallic. To replace it we generated the same lines with Cartesia, a text-to-speech service, but the pauses and word lengths differed from the video's speech: shifting the track in the edit left the lip sync off. We got a better result by converting the original voice with ElevenLabs Voice Changer, which preserved its rhythm. These are problems that surfaced in this particular production, compounded by the additional difficulties of our dubbing workflow. To reach the final edit we had to select alternative takes and work on the audio separately.