Higgsfield, reviewed

Higgsfield was founded in October 2023 by Alex Mashrabov, formerly head of generative AI at Snap, and Yerzat Dulat. It began as a video app for consumers before turning to creators and marketing teams.

On 29 September the company said it had passed 1 billion dollars in annualised revenue, counted as the last four weeks multiplied by 13, with more than 32 million users and nearly a million paying subscribers. In August it raised 400 million dollars at a valuation of 5.4 billion, led by DST Global. These are the company's own figures. Here is what it sells, who it suits, what it costs and where to be careful.

What it is

Higgsfield is less a model than a studio built around other people's models. For video it offers Seedance, Kling, Sora 2, Veo 3.1, Wan and Grok Imagine. For images it offers Google's Nano Banana family, GPT Image, Flux, Seedream and Ideogram, next to its own photo model Soul 2.0, which it describes as built for fashion and aesthetics. Forbes reported in February that Higgsfield had become the largest customer of OpenAI's Sora 2 by both spend and usage.

Around the models sits the part that is its own. A library of presets for camera moves and effects, which Forbes counted at 400. Studios for cinema scenes, ads, marketing, lip sync and fashion. Small apps for new angles, zooms, transitions and colour grading. A canvas for moodboards and chained workflows, an agent called Supercomputer, an API, and connectors that let you generate from inside Claude and ChatGPT.

Who it suits, and how good it is

The platform is built for volume. Higgsfield says direct to consumer brands drive much of its demand, because advertising is iterative: a first version, many variants, versions for other markets. For that work, one account with most major models and ready made camera moves saves real time, and a team can switch models per shot without new contracts.

Quality is mostly the quality of the model you pick, and those models are among the best available. What Higgsfield adds is direction in a box: a preset that produces a cinematic camera move without a word of camera language. In our view that is its strength and its limit. Presets make good looking clips easy, and they make many brands' clips look alike. The company itself writes that more output gives a brand more to test but “doesn't decide the outcome on its own”.

What it really costs

Everything runs on credits. In October the Starter plan cost 19 dollars a month with 270 credits, Plus 47 dollars a month billed annually or 59 monthly with 1'200 credits, and Ultra 99 dollars a month billed annually or 129 monthly with 3'000 credits. Team plans start at 65 dollars per seat. The free plan does not include commercial use.

The credit menu varies widely. A Nano Banana Pro image costs 2 credits, 4 at 4K. Five seconds of Kling 3.0 at 1080p costs about 8 credits, four seconds of Veo 3.1 at 1080p about 40, and four seconds of Sora 2 Pro at 1080p about 50. On Plus, billed annually, a credit works out at roughly 4 US cents, so that Sora clip costs about 2 dollars before any retakes.

The terms matter as much as the prices. Subscription credits do not roll over, credits bought as add-ons expire after 90 days, and a refund is possible within seven days of the first purchase only if no credits have been used.

Unlimited, and the question of trust

Higgsfield has promoted “unlimited” offers heavily, and much of the criticism it has faced concerns them. Users of a 25 dollar a month unlimited Black Friday plan told Forbes in February that processing slowed to a crawl after a few videos a day. The company blamed part of it on bot attacks, said it had shut down around 40'000 accounts and said it had refunded 1.35 million dollars to genuine users. In August, WION reported user complaints of waits of two hours and more on an unlimited Seedance offer that promised an average under five minutes; the company has said speeds can vary when demand peaks. Today's terms say unlimited use is subject to fair use limits and may be throttled or queued, and the pricing page says unlimited models do not work through its connectors and command line tool, Canvas or Supercomputer.

The marketing has drawn criticism of its own. Forbes reported that a media kit for paid creators contained stock video templates presented as AI, which the company said had been shared by mistake, and that promotional clips sent to creators included videos Forbes described as racist. A cofounder told Forbes: “It is absolutely not representative of our values, and it was a mistake.” The same month, The Register reported that Higgsfield had removed a post on X boasting that its AI motion design tool had put an end to more than 20 creative jobs, after criticism from artists, and that the company did not respond to requests for comment. The company has since said it grew its customer support team from two people to more than 40, and its chief executive, Alex Mashrabov, has acknowledged that “our internal processes and external communications haven't always kept pace”.

Lock in, and our verdict

The terms are fair on ownership: Higgsfield claims no ownership of what you make, does not restrict commercial use, and your rights survive cancellation. They are less comfortable on data. Unless a company signs an enterprise agreement, its inputs and outputs may be used to train Higgsfield's models. Deleting content stops that from then on, but content already used for training cannot be taken out again. The real lock in is softer: prompts, presets, canvases and habits that live inside one platform, while the models underneath are available directly from their makers.

Our verdict: Higgsfield is a fast way to try most major models and to produce ad variants at volume. Use a paid plan for anything commercial, treat “unlimited” as best effort, keep confidential product material out unless you have an enterprise agreement, and export everything you want to keep. Access to models is now easy to buy. A look that holds across hundreds of images is not, and that is what Reel is built for: styles locked from real references, the best model for each job, a quality check on every result and a person who picks.