AI Video API Pricing: What a Finished Clip Actually Costs

The first month I moved our ad renders off the web apps and onto APIs, I budgeted about $80 and spent a little over $300. Nothing had gone wrong. The per-second rate I'd planned around was exactly what the provider charged. I'd just never counted the clips I threw away.

That's the gap this article is about. Every pricing page in this space quotes you a number per second of output, and every one of those numbers is real and none of them is what you'll pay. So here are the current list rates, and then the four things that sit between that rate and your actual invoice.

The number that matters isn't per-second

Per-second pricing describes a render that succeeds, that you keep, and that needs nothing done to it afterwards. In practice we keep roughly one clip in three. Sometimes the hands are wrong, sometimes the product drifts halfway through, sometimes the model just ignores the camera instruction. All three of those bill exactly the same as a perfect take.

So the honest unit is cost per usable clip, and on our pipeline that runs two to four times the sticker rate depending on how fussy the shot is. A static product turntable lands near 1.5x. A person talking, moving, and holding something lands closer to 4x. We wrote about that ratio separately in what a finished AI video really costs.

The list rates, as they stand

These are published API rates rather than consumer subscription pricing, which is a different and usually worse deal per second. Figures below are from a July 2026 survey of provider pricing — treat them as a snapshot, because this is the fastest-moving number in the entire category.

Source for the rate column: buildmvpfast's API cost tracker, updated July 2026. We've cross-checked the Kling and Seedance numbers against our own invoices and they're in the right neighbourhood; the Veo tiers move more than the others.

The spread is the story: fifteen times between the cheapest and the most expensive paid option. That is not a quality gradient in any simple sense — it's mostly about whether audio ships in the same render.

Where the cost actually hides

Four things, in the order they've surprised me.

Retries. Already covered above, and it's the big one. Budget your discard rate before you budget your rate card. If you don't know your discard rate yet, assume two thirds and be pleasantly surprised.

Audio as a second pass. Veo 3.1 is currently the only model in the top tier that ships synchronised audio inside the video render. Everything else means a separate text-to-speech or music generation, which costs money and — more annoyingly — costs a round trip. That's why Veo's $0.75 isn't as outrageous as it looks next to Kling's $0.075: you're comparing a finished clip against a silent one.

Resolution upgrades. Most providers price 1080p at roughly four times 720p, not twice. If the clip is going to Instagram or TikTok, 720p is genuinely fine, and we render our whole social pipeline there. See aspect ratio and resolution for where that stops being true.

Failed jobs that still bill. Content-filter refusals are usually refunded. Jobs that complete but produce garbage are not. Those are two different failure modes and only one of them is free.

The honest limits of this comparison

I want to be straight about what a table like the one above can't tell you.

It can't tell you which model finishes your shot. Per-second cost is meaningless if a cheap model needs six attempts at a shot an expensive one nails first time — at which point the expensive model is cheaper. We've had that happen in both directions depending on the subject.

It also doesn't cover rate limits or queue times, and those cost real money when you're on a deadline. A provider at half the price with a twenty-minute queue is not half the price. And it ignores the commercial-use and data-residency terms, which matter enormously for client work and not at all for a personal project.

Per-model detail lives in the individual breakdowns: Veo 3, Kling 3, Seedance, Runway and Luma.

Cloud API versus running it yourself

There is a third option that the pricing tables never include, which is running an open-weights model on your own hardware. Wan is the obvious candidate. The marginal cost of a clip then is electricity, and the fixed cost is a GPU you either own or rent.

The crossover is further out than people expect. At roughly $0.05–0.09 per second on the cheap cloud tiers, you can render an awful lot of video before a serious GPU pays for itself — and you're doing your own ops the whole time. It makes sense when you're generating continuously, when you need the throughput, or when the content can't leave your machine. It does not make sense to save money on twenty clips a month. More on the tradeoff in open-source AI video generators and the Wan guide.

How I'd budget it now

If I were starting this month, with the benefit of that $300 first invoice:

• Prototype on a free or unlimited tier. Get the prompt right where iteration is free. We run our own social batch on Seedance 2.0 Mini through a web app's unlimited tier at zero marginal cost, and only reach for the API when the free path can't do the shot.
• Pick one model and learn it. The cost of switching models constantly is retries, and retries are the expensive part. Prompt fluency in one model beats a spreadsheet of eight.
• Multiply the rate card by three. Then decide if the project still makes sense. If it only works at 1x, it doesn't work.
• Batch overnight. Queueing a night's renders and reviewing in the morning removes the temptation to re-roll a clip four times while you watch it. See batch generation.

One footnote worth knowing: Sora's API was retired in September 2026, so if you find an old cost comparison that leads with it, the whole table is stale. We covered that in the Sora API shutdown.

Frequently asked questions

What's the cheapest AI video API right now?
On raw per-second rates, Veo 3.1 Lite and Wan 2.6 sit around $0.05/sec, with Kling 3.0 just above. Cheapest per usable clip depends entirely on how often the model gets your shot right first time.

Why is Veo 3.1 Standard so much more expensive?
It renders synchronised audio inside the same job and outputs at higher resolution. Compared like-for-like against a silent clip plus a separate voice pass, the gap narrows a lot — though it's still the premium option.

Do I get charged for failed generations?
Content-policy refusals are generally refunded. Jobs that complete and return something unusable are billed in full, and those are the majority of what you'll throw away.

Is the API cheaper than a subscription?
Usually, per second, if you use it heavily. Subscriptions win when your volume is low and predictable, because you're buying convenience and a UI as well as compute.

How many seconds should I budget per finished video?
Take your final runtime, multiply by three for discards, and add a little for the shot that always needs one more attempt. A 30-second ad realistically costs 90–120 seconds of generation.

Where to take this next

Pricing tables go stale within weeks in this category, which is exactly why I stopped relying on them and started tracking my own cost per usable clip instead. If you want to compare notes on what you're actually paying — and see the invoices and render logs behind the numbers above — that's the sort of thing we work through together every week.

Looking for something else? Browse all 140 AI video guides in one list.

Join the AI Video Generator community on Skool and bring your own numbers. Nothing sharpens a budget faster than seeing someone else's.