The setup
You wire it in once, then you ask Claude for a video the same way you ask it for anything else. No app to open, no timeline, no monthly fee — just a key on an account you top up when it runs low.
Read section 2 before you cancel anything. Pay-as-you-go is genuinely cheaper in bursts and genuinely more expensive if you generate constantly. I have put the real numbers there and a prompt that does the sums on your usage.
Open Claude Code and paste this. It picks a provider with you, walks you through getting the key, wires it up, and makes one test video so you know it works before you spend anything real.
I want to generate AI video by paying per generation instead of paying a monthly subscription. Set this up for me. I am not a developer, so explain each step in plain English and stop and wait whenever you need something from me. Do it in this order: 1. Ask me two questions first: roughly how many videos I make in a typical month, and whether I need audio in them. Do not skip this — the answers change what you recommend. 2. Based on my answers, recommend ONE provider from: fal.ai, Google's Gemini API direct, or Replicate. Tell me why that one, and what it costs per second or per video TODAY — go and read the provider's live pricing page, do not quote a price from memory. If the price you find is different from anything I have read elsewhere, tell me. 3. Walk me through getting an API key on that provider, and tell me exactly how much to put on the account to start. I want the smallest sensible amount, not a big top-up. 4. Store the key safely as an environment variable on this machine. Do not paste it into any file that could end up in git. Tell me plainly where you put it and how I would delete it later. 5. Write me one small script I can run that takes a text prompt and gives me back a video file, saved into a folder in my home directory with the date in the filename. Keep it simple and comment it so I can read it in six months. 6. Run it once with a cheap, short test — the cheapest model and the shortest duration the provider allows. Tell me exactly what that test cost before you run it, and show me the actual charge afterwards. 7. Finally, write me a one-page note in that folder covering: which provider, which model, what each video costs, how to check my balance, and how to top up. Assume I have forgotten all of this in three months.
Once that is done, generating is one sentence. This is the prompt I use, and the specifics in it are what stop you burning credits on a clip you cannot use.
Make me a video with the setup we built. Shot: [describe what is in frame, the camera move, the lighting and the mood — be specific, "a woman at a desk" gets you a stock photo that moves] Length: [seconds] Resolution: [720p or 1080p] Audio: [yes or no — audio usually doubles the price, so say no unless I need it] Before you generate: tell me what this exact clip will cost and what my balance is. Wait for me to say go. After you generate: save it with a filename I can search later, tell me the actual cost, and tell me the new balance. If the result is not usable, do NOT automatically try again — tell me what you would change in the prompt and let me decide whether it is worth another charge.
The bit people skip
That last instruction — do not automatically retry — is the difference between this being cheap and this being a surprise. An agent that quietly regenerates four times because the first three "were not quite right" has just spent your afternoon's budget while you made a coffee. Make it ask.
Do the sums before you cancel
The claim "the API is cheaper" is only half true, and the half that is false will cost you money. Here is the honest version.
Real per-video prices, read off fal.ai's live Veo 3.1 page on 08/09/2026. Their words: "For every second of video you generate you will be charged $0.20 without audio or $0.40 with audio for 720p or 1080p." All figures USD.
An 8-second 1080p clip, pay-as-you-go
- Veo 3.1 Lite, no audio — about $0.40
- Veo 3.1 Fast, no audio — $0.80
- Veo 3.1 Fast, with audio — $1.20
- Veo 3.1 full quality, with audio — $3.20
No minimum, no monthly fee. You pay when you generate and nothing when you do not.
A subscription, for comparison
- Higgsfield Starter — $19/month
- Higgsfield Plus — $59/month, or $47 on annual
- Higgsfield Ultra — $129/month, or $99 on annual
Read off their pricing page on 08/09/2026. There is a free tier too. Their plan names and prices have changed more than once this year, so check them yourself.
Divide one by the other and you get the line. Against the $59 plan, at Fast-with-audio prices, pay-as-you-go wins until you are making about 49 videos a month. Against the $19 plan, the line is about 16 a month. At full-quality prices those numbers drop to roughly 18 and 6.
So, roughly
Under about 15 videos a month: pay-as-you-go, comfortably. It is not close.
Somewhere between 15 and 50: it depends on your plan and which model you use. Do the actual sum.
Well over 50 a month, every month: stay on the subscription. You are the customer it is priced for, and you will lose money switching.
But the volume is not really the point, and this is the part worth understanding:
Subscription credits expire. Prepaid credits do not.
Higgsfield's own FAQ, word for word: "Unused subscription credits do not roll over to the next billing cycle and expire at the end of each month." That is completely normal — nearly every credit subscription works this way. It just means a quiet month is money you do not get back.
Three quiet months on the $59 plan is $177 for nothing. On a prepaid balance, three quiet months costs you nothing at all, and the credits are still sitting there when you come back.
So the real question is not "which is cheaper". It is are you paying for capacity or for output? If your work comes in bursts — a launch, a client project, then three weeks of nothing — you are buying capacity you do not use.
Run this on your own numbers before you touch your billing settings. It goes and reads today's prices rather than trusting mine.
Work out whether I should move off my AI video subscription and pay per generation instead. Be honest with me even if the answer is that I should stay. My situation: - Tool I pay for: [name] - What I pay: [$X per month, or $Y per year] - Videos I actually made in each of the last three months: [e.g. 40, 0, 6 — go and count, do not estimate upward] - Typical clip: [length in seconds], [720p or 1080p], audio [on/off] Do this: 1. Go and read the CURRENT published price per second or per generation for an equivalent model on fal.ai, Google's Gemini API and Replicate. Use today's live pages. Give me the figure, the date and the URL for each. If a price is not published without logging in, say so rather than guessing it. 2. Work out what my last three months would have cost on each of those, versus what I actually paid. Show the arithmetic so I can check it. 3. Work out my break-even: how many videos a month before the subscription becomes the cheaper option. 4. Now argue the other side properly. What does my subscription give me that an API does not — a interface, presets, models the APIs do not carry, unlimited-use periods, support? Put a rough dollar value on the ones I would actually miss. 5. Give me a straight recommendation: switch, stay, or run both. If the saving is under about $20 a month, say so and tell me it is not worth the hassle. 6. If the answer is switch, tell me what to do about anything I have already paid for — an annual plan I am mid-way through, credits sitting unused, work in progress inside the app.
One thing that will skew your answer
Count your last three months honestly, including the quiet ones. Everyone remembers the launch week where they made forty clips and forgets the six weeks either side. The quiet months are the entire argument, and they are the ones you will leave out.
Where to actually buy the generations
Four routes, all real, all checked on 08/09/2026. They are not equivalent, and one of them I would not build a business on — which I have said in its own card rather than quietly leaving it off.
fal.ai
Where I would startNo minimumHosts the major video models and publishes every price openly, no login required. Their own words on the Veo 3.1 page: "Pay-per-second with no minimums." Veo 3.1 runs $0.20 a second without audio and $0.40 with, at 720p or 1080p; the Fast tier is $0.10 and $0.15. They also do the worked example for you — a five-second 1080p clip with audio is $2.00.
What it will not do: hold your hand. There is no timeline, no preset library and no community. It also has no official video-generating connector for Claude — their MCP server is a documentation search tool, not a generator — so the setup prompt above has Claude write you a small script instead. That works fine and it is one file, but it is a script, not an app.
Google's Gemini API, direct
Straight from the sourceIf the model you want is Veo, you can buy it from the people who made it. Standard Veo 3.1 is $0.40 a second at 720p and 1080p with audio included, $0.60 at 4K; the Fast tier is $0.10 to $0.12, and Lite is $0.05 to $0.08. Their billing note is the one that matters: "You will only be charged if your video is successfully generated."
What it will not do: give you anything except Google's models. No Kling, no Runway, no Seedance — if you want to compare models, this is the wrong door. There is also no free tier for Veo, so the first thing you do is attach a card.
Runway and Replicate
No subscriptionBoth sell generation as pay-per-use with no monthly commitment. Runway's API sells credits at a flat cent each and lists what each model costs per second — Veo 3.1 with audio at 40 credits a second, its own Gen-4 Turbo at 5. Replicate charges by what the model actually uses and has no minimum spend, with the price on each model's own page.
What it will not do: quote you a simple headline price. Replicate in particular bills a lot of models by compute time rather than per video, so two runs of the same model can cost different amounts. Fine once you are comfortable; confusing as your first stop.
Kie.ai
Cheapest, and the most caveatsA reseller. It buys access to other people's models in bulk and undercuts them, which is why it turns up in every "cheaper than the official API" video. The savings are real and the arithmetic checks out: it lists Veo 3.1 at 1080p for around $1.16 to $1.28 a video against an official $3.20, and Google, fal.ai and Runway all independently price that model at $0.40 a second, so the discount is genuine rather than a marketing number. Credits are $0.005 each with a $5 minimum, and "credits never expire". It is also the one route with a real Claude connector — a community-built one, @felores/kie-ai-mcp-server, MIT-licensed and last updated 23/08/2026:
claude mcp add kie-ai --env KIE_AI_API_KEY=your_key -- npx -y @felores/kie-ai-mcp-server
What it will not do: reassure you. Read this bit properly before you put money on it. It has no terms of service page at all — which means no published refund policy and nothing written down about commercial use of what you generate. It sits at 2.2 out of 5 on Trustpilot across a small number of reviews, and the recurring complaint is credits disappearing on generations that failed. Anything you make is deleted after 14 days, so download as you go. It is not a licensed partner of most of the models it resells, and Midjourney has already had it removed. On 08/09/2026 its own homepage and its own Veo 3.1 model page quoted two different prices for the same thing. Also worth knowing: it does not carry Sora at all. Cheap enough to experiment with, and I would not run client work through it.
What you give up
This is a real trade, not a free upgrade. Four things you lose, said straight.
A polished app, for a key and a script
The first setup is fiddlier than opening a website. You get an API key, a balance and a folder of files — not a library, not a preview grid, not an undo button. The setup prompt above makes that afternoon short, but it does not make it zero.
The presets, and the taste baked into them
This is the one people underestimate. Those camera moves and viral presets are somebody's craft, packaged. Go to the API and you are writing every shot from scratch, and your first ten will be worse than what the app gave you. Budget for that in credits, not just in time.
Some models entirely
Subscription apps license models the resellers do not have. Sora is the obvious one — it is on Higgsfield and it is not on Kie.ai. If a specific model is the reason you pay, check it is reachable before you cancel, because "I will just use the API" does not help if the API does not sell it.
Unlimited, where it exists
Some plans include unlimited-use windows on named models. Nothing priced per generation can compete with unlimited — the marginal cost is zero and yours never is. If you genuinely use those windows, the subscription is doing something for you that no amount of arithmetic will beat.
And the thing you take on
A balance you have to keep topped up. It is a small thing until it is 11pm before a launch and the generation fails because the account hit zero. Set an auto top-up, or put a recurring reminder in your calendar to check it. That is the whole maintenance story.
The version I would actually recommend
Do not cancel anything this week. Put $10 on one of the routes above, run the setup prompt, and make five videos the API way while your subscription is still running. You will find out in an afternoon whether the loss of the presets bothers you, and you will have real numbers instead of my estimates. Cancel at the end of the month if it worked, and if it did not, you spent ten dollars finding out.
Prices move. These were checked on 08/09/2026
Everything on this page came off a live pricing page on the day it was written, and video pricing changes faster than almost anything else in AI — two of the tools here renamed their plans more than once this year, and one of them was quoting two different prices for the same model on the same day. So use my numbers to see the shape of the decision, and use prompt 3 to get today's real ones before you act on it. If you find something on this page has gone stale, tell me and I will fix it.
Ready to go deeper?
Connectors are the easy half. The useful half is knowing what to point them at.
Wright Mode Membership
Join a community of women using AI to take real work off their plate. Live calls, copy-paste prompts and workflows, and ongoing support.
Claude Masterclass
Learn to use Claude like a pro — from prompting fundamentals to building real workflows that save hours every week.
Claude Code Masterclass
Go beyond the chat box. Build automations, process data and create tools with Claude Code — no developer experience needed.