Higgsfield API: Setup, Per-Second Prices and Billing Rules

Yes, Higgsfield has a public API, and it is a separate product from the plans on higgsfield.ai. Since September 2026, open.higgsfield.ai has sold one key for 50+ image, video and audio models, paid from a prepaid US-dollar balance with a $5 minimum top-up, charged per generation, with no subscription and no seats.^[1] No Higgsfield plan includes it, and plan credits and Unlimited windows do not apply. Our verdict is simple. Pick the API to wire generation into your own app or pipeline; for hands-on work in the browser that uses most of a plan's monthly credits, a paid plan costs less per clip, while light or irregular use can cost less on the API.

Separate key, separate balance, separate rules

The API sits apart from your Higgsfield account. Subscription credits cannot pay for API calls, and the API Terms state that consumer, Team and Scale subscriptions carry no API entitlement; usage draws on its own prepaid balance.^[2] The confusion is understandable, because Higgsfield also offers MCP, a CLI and plugins that look like developer tools. They are not the API. MCP, CLI and plugins sign in with your Higgsfield account and spend plan credits; the API uses its own key and a separate prepaid dollar balance. To drive Higgsfield from a Claude chat, read our Higgsfield MCP guide instead.

The catalog is not the website lineup either. The docs list video families from Seedance, Kling, Wan, Happy Horse, Cinema Studio, MiniMax, LTX, PixVerse, Grok and Genjutsu, plus image models such as Soul 2, Marketing Studio, Ads Studio, Recraft, Qwen Image and Ideogram. Veo is missing from the API in November 2026, even though the website offers Veo 3.1 from the Plus plan. The API also excludes the Supercomputer agent, GPU compute and text or LLM models. Higgsfield says model access can change, so check the console for your account. Verify before you build.

From sign-up to a finished clip in six steps

Every model shares one authentication scheme and one asynchronous request lifecycle; only the JSON body changes from model to model. That keeps the learning cost low once the first integration works. The base URL is api.higgsfield.ai, and every call carries an Authorization header in the form Key KEY_ID:KEY_SECRET. Keys must stay on your server: the Terms forbid embedding them in client-side or public code, and a server-side proxy pattern is documented for browser and mobile apps. Keys belong to a workspace rather than a person, and in organizations a key keeps working after the member who created it leaves.

1. Open the console at open.higgsfield.ai and create a key. You get a key ID and a secret, and the secret is shown only once.
2. Add funds to the balance; the minimum top-up is $5, and auto top-up can refill it.
3. Price the job first with the estimate endpoint (POST /estimate/ plus the model path), which returns the cost in US dollars.
4. Submit JSON to the model endpoint with prompt, resolution, duration and aspect ratio, add an Idempotency-Key, and store the request_id, status_url and cancel_url you get back.
5. Poll the status_url, starting at two seconds and stretching to ten, or pass a webhook in the hf_webhook parameter.
6. Download the output once the status reads completed, before the time given in expires_at.

Official SDKs cover Python (higgsfield-client) and TypeScript (@higgsfield/client). They handle authentication, submission, polling, cancellation and file uploads, and they run server-side only; other languages use plain REST. You can cancel only while a job is still queued. For production, Higgsfield recommends webhooks over polling. Endpoints need public HTTPS, a 2xx reply within ten seconds and signature checks under the Standard Webhooks scheme. Failed deliveries retry for up to two hours. Duplicates can arrive.

What a second of video costs at list price

Video models are priced per second of output, image models per image, and DoP per generation. The figures below are list prices from the API price page and the model playgrounds, read in November 2026.^[3] Many rates shift with resolution, input type and aspect ratio, and video references can bill both input and generated seconds. For your exact request, the estimate endpoint has the final word, so treat this table as a planning guide rather than a quote.

Treat every discount as an offer, not a price. In November 2026 the API pages advertised a launch offer of 15% off all models with up to 75% off three models you choose, a $15 balance after you verify a business email and add a card, max-discount prices 30% to 50% below list, and 100% cashback banners without terms. A separate cashback program pays $15 after your first $100 of spend and up to 20% back by 30-day spend tier. The Terms add a catch: discounts may not apply to partner models run by third parties. Budget at list price.

Billing rules that change the real bill

The API charges only for work that finished. Failed, nsfw and timed-out requests cost nothing, and the reserved amount returns to the balance automatically; a canceled queued request is refunded too.^[4] Server errors (5xx) and invalid-input rejections (422) are not billed. A usable clip you reject is billed, so a weak prompt still costs money. Funds expire one year after you add them and are non-refundable except where law requires. The balance cannot go below zero, so requests fail at zero unless auto top-up is on. Invoice billing needs a committed-use contract.

Concurrency is the main limit, and it is tied to money. A new account runs 2 generations at once. Funds added in the past 28 days raise that to 10 at $25 or more, 20 at $100 or more and 40 at $1,000 or more; anything higher goes through sales. Requests above the limit wait in a queue and start on their own. When the queue fills, you get a 429 with Retry-After, and nothing is created or charged. Some models carry their own separate limit, and rate-limit headers do not exist yet. The practical rule: if a pipeline needs ten parallel jobs, keep at least $25 of top-ups inside every rolling 28 days.

When a Higgsfield plan beats the API per clip

For steady volume, plan credits win. By our arithmetic at list price, a 5-second Kling 3.0 Standard clip through the Higgsfield API costs $0.42, while a 5-second Kling 3.0 clip at 720p on a Higgsfield plan uses about $0.34 of Plus monthly credits or $0.23 of Ultra yearly credits. At Higgsfield's posted credit rate, a 5-second Seedance 2.5 clip at 720p comes to about $1.72 of Plus monthly credits or $1.16 of Ultra yearly credits, against about $2.31 on the Higgsfield API. Kling 4.0 at 1080p follows the same pattern: about $1.18 a clip through the Higgsfield API, $0.98 on Plus monthly and $0.66 on Ultra yearly.

Those gaps hold only if you spend the credits. Plan credits reset every month (every 30 days on yearly billing) and unused ones are lost, which our Higgsfield pricing guide covers. API funds last a year. So the threshold is usage, not price. If you would leave much of a plan's monthly allowance unspent, the API's higher per-clip rate costs less than credits that expire. Retakes hit both sides equally, and Higgsfield's own post says three to five generations per usable clip is normal, so budget several takes either way.

The API also wins on things a plan cannot do. Unlimited windows never cover automated use, and scripting is banned on them, so any pipeline pays per generation anyway. The API adds webhooks, idempotent retries, a cost estimate per request and concurrency you can buy. A plan wins for hands-on creators who use most of its monthly credits, for Veo 3.1, and for Cinema Studio above 720p: the API runs Cinema Studio 4.0 at 480p or 720p only, while the website goes to 1080p.

Data, consent and the limits of a self-serve account

Output files stay available for at least 7 days and may be deleted after that, so copy them to your own storage. Higgsfield can set account retention between 1 hour and 7 days on request, the Hf-Expires-After header can only shorten it, and a DELETE call removes files, prompt and parameters at once while billing records stay. Zero data retention exists only under an enterprise addendum. API content may train Higgsfield models unless your workspace opts out in the dashboard. The opt-out is off by default, so training applies until you switch it on; the change takes up to 10 business days and is not retroactive.

The rules for people are strict. Face swap, lip sync, avatar and Soul ID endpoints may require identity or business verification, and any real person's face or voice needs documented consent covering both the input and how you use the output. Your end users must be 18 or older, with age-gating on your side. Facial recognition is banned, and so is reselling the API as a standalone generation service. Higgsfield claims no ownership of outputs but gives no warranty that they are non-infringing. Self-serve accounts get no uptime commitment; SLAs need an Order Form. Technical help runs through Discord.

The bottom line

The Higgsfield API is a clean developer product with honest billing rules: a $5 entry, per-second list prices, no charge for failures, funds that last a year and concurrency that grows with what you add. It is not a perk of any plan, and it is not the cheapest route per clip for anyone who uses most of a plan's monthly credits by hand. Use it when generation has to run inside your own software, when volume is bursty or light, or when you need parallel jobs on demand. If you would spend most of a plan's credits every month, a plan's credits stretch further.