Plug AI image and video generation into your own tool or client portal

This page is written for the person who will brief a developer, not for the developer. The word "API" turns up in a lot of searches for people who want code samples. If that is you, the vendor documentation is better than I am. If you run an agency or a business and need to decide whether to pay for an integration, which vendor to pick and what it will really cost, read on.

I cover the cheapest options first, what an API actually does, the main vendors with prices from their own pages, where the Higgsfield API fits and where it does not, and a pilot you can run before you commit. Prices were read on 3 November 2026 and change often, so confirm on the live pages.

This page may contain affiliate links from which Blur Studio earns commissions.

When copying and pasting stops working

Manual work is fine for a long time. It stops being fine when:

• the same task repeats every week for many clients or products,
• people lose track of which file is the approved one,
• clients should request or approve assets themselves, or
• you need to know what each client's assets cost you.

A useful rule of thumb, which is mine and not a published benchmark: under about 20 assets a month, use a website or a chat tool. Between about 20 and 200, a spreadsheet plus a simple script and manual approval is usually enough. Above about 200, or when clients need to generate for themselves, a real integration starts to pay off.

Cheapest first: spreadsheet, manual uploads, then chat

1. A spreadsheet and manual uploads. One row per asset: client, brief, status, link to the file, approved or not. It costs nothing and shows you where the pain is.
2. The tool's website. Most generation platforms have a project or folder system. Try it before you build.
3. A chat assistant with a connector. See making ads by chat in Claude or ChatGPT and the Higgsfield MCP guide. That avoids building anything, but it has no hard spending cap and is not an approval trail.

If you still have a problem after these three, then look at an API.

What a generation API actually does

Using one is a three step loop:

1. Submit a request with a prompt, a model name and any reference files. The service replies with a request ID straight away.
2. Wait. Generation takes time, so the work is asynchronous. Your program either checks a status link every few seconds (polling) or the service calls your program when it is done (a webhook).
3. Download the result and save it somewhere you control.

Everything else, such as approvals, client folders, spend tracking, error messages and storage, you build around this loop. That is why an API is a project and not a setting. A developer can build a first version in days, but the rules around it are the real work.

Do you need a generative API or a template API?

Search results for "video generation API" mix two different products. Orshot, which sells a template renderer and so has a stake, puts it this way: a generative API invents footage from a prompt, different every run and billed per second, while a template API fills a design you made, the same every run. Its test is that if the output must look identical every time, such as a logo in the same corner of 500 videos, you want a template tool. If it should look new every time, you want a generative one.

Motionbox, another video tool vendor, adds a practical point: model APIs return short, silent clips, usually 4 to 12 seconds, with no captions or brand assets, so most teams still need a step that assembles, captions and brands them. Plan for that step in your budget. For product catalogues, automatic product videos compares both kinds.

The main options and what each charges

Vendor	What it is	Pricing shown (3 November 2026)	Good when
Runway API	The developer interface to Runway's video models.	$0.01 per credit. Gen-4.5 is 12 credits a second, so about $0.12 a second. Gen-4 Turbo is 5 credits a second. Runway's docs also list Seedance 2.5 at 30 credits a second (720p) and 68 (1080p).	You want Runway models and a clear per second price.
fal	A marketplace that gives developers one interface to many models.	Examples: Kling v3 Pro about $0.14 a second, Kling 2.5 Turbo Pro $0.07 a second, MiniMax H3 Max $0.03 a second, Nano Banana 2 $0.08 per image, Nano Banana Pro $0.15 per image.	You want the biggest model choice and standard developer tools.
Replicate	A marketplace for running open and commercial models through an API.	Flux 1.1 pro $0.04 per image. Wan 2.1 at 480p $0.09 a second, at 720p $0.25 a second. Other models are billed by computing time.	You need a flexible catalogue and can estimate machine time.
Higgsfield API	A pay per use interface to the models on Higgsfield, with its own dollar balance.	Per second and per image prices by model. I am not printing per model prices, see below. The estimate endpoint gives the cost before each run.	You already use Higgsfield and want its model range and tools in your own system.

Sizes differ, which makes comparisons hard. Kling, Seedance and others also appear on more than one vendor at different prices and different launch discounts. The only fair comparison is the one in the cost section below, using your own prompts.

A price gap worth knowing about

Runway's API credits cost $0.01 each. Its Standard web plan costs $15 a month for 625 credits, which is about $0.024 a credit, or $0.019 on the yearly price of $12. Gen-4.5 uses 12 credits a second either way. So the API works out to about $0.12 a second against about $0.29 a second on a Standard monthly plan. This is arithmetic from Runway's own pricing pages, and it shows that the web plan and the API can be priced very differently. Ask this question of any vendor.

Where the Higgsfield API fits and where it does not

From Higgsfield's help center and docs:

• It is a separate product with a separate balance. The API uses a US dollar balance, not your plan credits. Your plan, its credits and Unlimited models exist only on the website. A plan with Unlimited does not apply to the API.
• Minimum top-up is $5. Funds expire one year after you add them. The balance cannot go below zero, and auto top-up is available.
• 50 or more image and video models. Creating a key gives you 20 concurrent requests (requests running at the same time). The help center does not say whether more can be requested.
• Outputs are kept for at least 7 days. After that, assume they are gone. Your system must download and store every file itself.
• Failed requests are not billed. An estimate endpoint returns the price before you run a request.
• Audio is not listed. The help center names video and image models only. If you need voice or music, check the catalogue, and plan for a second source if it is missing. Runway's API pricing page lists audio models, including ones from ElevenLabs (a voice AI company).
• Commercial use is allowed under the Terms. Section 4.4 says Higgsfield does not restrict commercial use of Outputs. That does not clear a person's likeness. If a client uploads a photo of a real person, you need a record of that person's consent.
• Tools for developers. Python and TypeScript libraries, and a console where Analytics shows spend and request volume over time and exports to CSV. You can sign up as an individual or as a company.

I do not print Higgsfield's per model prices on purpose. The catalogue is a script rendered page that could not be read as plain text, and independent snapshots disagree because of launch discounts. Use the estimate endpoint on your own prompts, and read Higgsfield pricing explained for the website plans. I also do not claim the Higgsfield API is cheaper than the others. Nothing in the sources supports that.

The Higgsfield API is a good fit when your team already uses Higgsfield, you want its models and tools such as motion transfer inside your own system, and you can handle a developer and storage. The offer below is for Higgsfield plans on its website. The API runs on a separate balance, so check on screen what the offer covers.

Cost per approved asset: the one calculation that matters

Price per second tells you what you pay to generate. It does not tell you what you pay for something you can use. The calculation is:

cost per approved asset = (cost per attempt x attempts) divided by approved assets

An illustrative example, with made up acceptance rates. An 8 second clip on Runway Gen-4.5 is 96 credits, which is $0.96 at API rates. If you try 10 prompts, spend $9.60, and 3 clips are good enough to send, each approved clip costs $3.20. If 7 are good enough, it costs about $1.37. The price per second did not change, but the cost per approved clip more than halved.

So the one number to find out before you choose is your approval rate on 10 real prompts for your kind of work. A cheaper model with a 20% approval rate is more expensive than a dearer model at 70%. Retries after a timeout can also bill twice with some vendors, according to Magic Hour, so ask each vendor how it handles that.

Storage, retention and moderation you must plan for

• Retention. Higgsfield keeps outputs for at least 7 days. Magic Hour, a rival with a stake in the comparison, reports (sources checked 19 October 2026) that Replicate removes API inputs and outputs after one hour by default and that Runway result links expire within 24 to 48 hours. I did not check those figures against Replicate's or Runway's own docs, so confirm them. The safe design is the same for all: download within the job callback, store in your own bucket, and save the request ID with the client record.
• Moderation. API outputs fall under the same filters for copyright, brand names and unsafe content. Requests will sometimes be blocked or refunded. Your portal needs a plain message for that, not a spinner that never ends.
• Consent. If a client uploads a person's photo or voice, keep a record of that person's written consent, and check local law and the platform's current rules.
• Labels. Some ad platforms require AI labels. Build a field for it.

Terms to read before you build a client portal

Higgsfield's Terms of Use have a section for programmatic access (Section 11 in the live text I read). Four points matter for an agency:

• No pass-through. Section 11.5 says you may not resell or make access available on a standalone basis, or act as a pass-through or service bureau with no independent value added. A portal with approvals, storage and your own workflow adds value. A bare "type a prompt, get a video" reseller page is the risk. Ask Higgsfield in writing if you are unsure.
• Keep keys on your server. Section 11.3 forbids embedding keys in client-side or publicly accessible code.
• Do not call it human-made. Section 11.7 says you must not represent Outputs as human-generated and you are responsible for AI disclosures required by law.
• Training. The general content licence in Section 4.4 lets Higgsfield use inputs and outputs to improve its models, except under Enterprise agreements. Tell clients before their confidential material goes in.

A pilot with one client: 10 to 14 days

This is a plan you can run, with a rule for when to stop.

1. Pick one client and one asset type, for example 9:16 product clips.
2. Fund the minimum balance. For Higgsfield that is $5. For others, the smallest top-up they allow.
3. Write 10 real prompts. Run each through the estimate endpoint and record the price.
4. Run all 10. Record the wait time for each, and which are approved by the person who would sign off in real life.
5. On day 8 or later, try to open every file again.
6. Work out the cost per approved asset and compare it with what you pay now.

Pass only if all four are true: at least 7 of 10 approved, the median wait is shorter than your client will tolerate (measure it, do not trust any vendor's speed claim), no file is lost after day 8, and the cost per approved asset is below your current cost. If any one fails, stop. Then try a different model or vendor, or go back to the chat route.

What a build looks like

A first version is small and the rules around it are the work:

• A request form for each client with the brief, format and the claim they can prove.
• A queue that calls the API, keeps the request ID and handles waiting.
• Storage in your own bucket, with files named by client and version.
• An approval step before anything is released. See a client approval workflow and agency client portal software.
• Spend by client. The help page describes spend and request volume over time, and does not say it can be split by your own clients. Tag every request with your own client ID in your database. A separate page covers tracking AI spend by client.
• A fallback. If the API is down or a request is blocked, the portal shows a clear message and the team knows what to do.

If you do not have a developer, an automation tool such as Zapier or Make can call a generation API with a web request, so a careful non-developer can pilot with one scenario. Be careful with retries and timeouts, since a retry can bill twice. For catalogue scale video, see video for every product.

Who should skip the API

• You make only a few assets a month. The time to build and maintain an integration will cost more than you save.
• You have no developer or build partner and no one to look after it. An API is not a no code setting.
• Your clients want to review and revise in person. A portal does not help if the conversation happens on calls.
• You need real footage for a claim. Keep real footage where a claim must be real.
• You need voice or music from the same vendor and you are looking at Higgsfield. It has no standalone audio models, so plan a second source.
• You only want drafts in chat. The chat route needs no build at all.

For a wider tool comparison, see the best AI video generators for agencies.

How I'd approach this

I would not start with an API. I would run the spreadsheet and the website for a month, count how many assets I really make, and find out my approval rate. If the count is high and the pain is real, I would run the 10 to 14 day pilot with one client and one asset type, then decide. If I was already using Higgsfield, I would try its API first, with the estimate endpoint on every batch, and keep Runway or fal as a comparison.

If the pilot passes, the build is mostly the unglamorous parts: approvals, storage, spend by client and clear errors. Blur Studio builds that kind of repeatable creative workflow, client approval portal or shared asset library, so your team uses one screen and the vendor behind it can change later.