What Video APIs Cost per Finished Clip

Published examples include $0.03 per generated second at 720p on xAI, $0.19 for a six-second 768p MiniMax clip, and 100 Luma credits for a five-second standard 720p request. Those numbers cannot be compared until you choose a model, resolution, audio setting and shot length. Your finished-clip bill also includes rejected generations and footage you cut. For example, three accepted six-second MiniMax Hailuo 2.3 Fast clips at its published 768p price of $0.19 each would cost $0.57 before other work; needing twice as many paid attempts would make that $1.14. The second figure is an illustrative retake scenario, not a measured model acceptance rate. The MiniMax price list supplies the unit price.

Compare ten API buying routes by their actual unit

The table names ten billing and access surfaces, not ten models of equal quality. Google Cloud and the Gemini API have separate pricing pages; Runway, Higgsfield, fal and Replicate are separate places to buy model access. A stated rate applies only to the named setting. “Not verified” means the official page opened for this article did not establish that detail; it is not a zero-dollar charge. Prices are public USD figures read on November 6, 2026, not account quotes.

API route and example	Published unit and example price	What to check before budgeting
Runway Dev, gen4.5	12 credits per output second; credits sell for $0.01, so $0.12/second	Model, requested duration and any separate input-media charge. Runway pricing
Google Cloud, Veo 3.1 Fast	$0.10 per generated second at 720p with audio; $0.08 without audio	1080p and 4K rates differ. Confirm region and the model’s supported lengths. Google Cloud pricing
Gemini API, gemini-omni-1.1-flash	$17.50 per million video-output tokens; Google says 5,792 tokens per 720p second, about $0.10/second	Text and multimodal inputs have a separate $1.50/million-token rate. Text output is $9.00 per million tokens. Do not treat the approximate output figure as the total invoice. Gemini pricing
BytePlus ModelArk, Dreamina Seedance 2.5	Official estimated example: $1.156 for a five-second 720p, 16:9 output without input video	Token use, aspect ratio and input video can change the bill; this is an illustrative billing example, not a fixed five-second tariff. BytePlus pricing
MiniMax direct API, MiniMax-Hailuo-2.3-Fast	$0.19 per 768p six-second video; $0.32 per 768p ten-second video	These are fixed clip combinations, not a linear per-second rate. H3 models have different rates and input charges. MiniMax pricing
Luma Agents API, ray-3.2	Standard 720p video generation: 100 credits for five seconds or 300 for ten	Credit-to-USD value depends on account plan; edit, extend, reframe and HDR use other tables. Luma pricing
Higgsfield API, Kling 3.0 route	Page shows a sale rate of $0.042 per video second against $0.084 regular	Promotion and selected model matter; do not assume the sale rate lasts or equals Kling’s direct price. Higgsfield API
fal, Kling 2.5 Turbo Pro route	$0.07 per generated second in fal’s public video listing	Pricing varies by model; credits are prepaid, and errors and queued time have separate rules. fal video listing; billing rules
Replicate, google/veo-3.1-fast	$0.15 per output second with audio or $0.10 without audio	These are Replicate’s own endpoint rates, not Google Cloud’s; inspect the exact model page before ordering. Replicate model page
xAI direct API, grok-imagine-video-1.5-lite	$0.03 per output second at 720p; $0.01 per input image	The 480p and 1080p rates differ, and image input adds a charge. xAI pricing

The table is a purchasing map, not a cheapest-to-best ranking. Google Cloud’s Veo Fast row and Replicate’s Veo Fast row illustrate why a provider name alone is not a price: the reseller’s audio-on rate is different from the cloud rate. Equally, Luma’s 100 credits have no universal dollar value without the account conversion. BytePlus labels these prices as estimates for specified inputs and outputs; use the API usage field to check actual token consumption.

Google’s Gemini Omni Flash is token-billed even though the public page provides an approximate 720p per-second equivalent. The arithmetic behind that approximation is 5,792 output tokens × $17.50 / 1,000,000, or about $0.101 per output second. Add billed input tokens and any text-output tokens separately; text output is $9.00 per million tokens. That makes it unsuitable for a spreadsheet column that assumes every listed “$0.10/second” is the entire request price.

Turn a shot list into billable generations

Start with the edited result you need, then map each shot to a supported generation length. Suppose a 15-second finished clip uses three five-second shots. If an API only sells six-second generations for the selected model, three accepted requests buy 18 generated seconds, even though the edit retains 15. The unused three seconds are a production choice, not a refundable balance. Conversely, a service that lets you buy a five-second output may fit the same shot list more closely. Do not interpolate a six-second fixed price into an imaginary five-second request.

Write one row per shot: required retained seconds, allowed generation duration, output setting, price for one successful request, and any reference input. A simple flat-request worksheet is attempts × price per request + separately billed inputs. For a per-output-second route, use billable output seconds × rate, then add input charges. For credit plans, keep the result in credits until you have your account’s dollar conversion. A token-billed video call uses output tokens and any billable input tokens, not just the video’s displayed duration.

References change the bill. xAI lists $0.01 per input image for video 1.5 Lite. MiniMax H3 charges for input-video duration and bills images beyond its free per-task allowance. Luma prices edits and extensions separately. Runway lists output-plus-input-video formulas for some models. A reference-heavy workflow can reverse a comparison based on output rates alone.

Our image API cost guide explains accepted still-image costs, while the text-to-speech API guide handles separately synthesized narration. Those charges belong in separate rows if your finished clip needs reference artwork or externally generated voice. The site’s AI API cost calculator helps with text-token planning; its model menu does not quote video-second or video-credit costs.

Budget the same clip at three acceptance rates

Use a hypothetical acceptance fraction to see how retakes affect the budget. It is the share of paid successful generations you keep for your edit, not a claim about a provider’s quality. You must measure that fraction on your own prompts before treating it as a forecast. Failed API jobs may be billed differently from completed videos you reject, so keep those in separate columns.

Take the three-shot, six-second MiniMax Hailuo 2.3 Fast example above. Each 768p six-second generation is $0.19 on MiniMax’s published legacy table. Three accepted clips retain five seconds apiece. Assume every paid request produces a clip, the only cost is the listed output fee, and the editor rejects whole clips rather than trimming or extending them:

Hypothetical accepted fraction	Paid attempts for three accepted shots	Output subtotal	Generated seconds; retained seconds
100%	3	$0.57	18; 15
50%	6	$1.14	36; 15
25%	12	$2.28	72; 15

The formula is paid attempts = accepted shots / accepted fraction, rounded up to a whole request, then subtotal = attempts × $0.19. These fractions are sensitivity cases, not test results. They exclude taxes, editing, storage, input media and any separate narration. If the actual acceptance fraction differs by shot type, calculate attempts for each shot rather than applying one global fraction. A first-shot prompt may need fewer retakes than a continuity-sensitive ending, but that is something to measure, not assume.

At 50% acceptance, 36 generated seconds are paid for even though the edit retains 15. Record generated output, accepted source and retained edit seconds separately.

Check failures, audio and unused credit

Do not put an automatic refund into the worksheet. The fal billing rules say server errors and queue time are not billed and that successfully generated outputs are charged. MiniMax’s package rules say failed generations or videos that trigger security review do not deduct package points. Those statements apply to their respective products and do not make a user-rejected successful clip free. Keep a separate field for each provider’s actual failed-job treatment.

Audio is another price switch. Google Cloud lists different Veo 3.1 Fast 720p rates with audio ($0.10/second) and without ($0.08/second). Replicate lists its Veo 3.1 Fast endpoint at $0.15 with audio and $0.10 without. A silent draft and a voiced final render are different line items. If you buy music or narration through another API, add that service’s meter rather than silently treating it as included.

Prepaid balances and promotions change cash outlay without changing the underlying output count. Luma’s credit value depends on the account plan, while Higgsfield’s displayed Kling rate is a sale. Check minimum top-up, credit expiry, tax and contract terms in your account before making a purchase decision. A low per-second rate is not a full production quote when your workload also needs edits, extensions, video references or a higher resolution.

Choose a route for the workload

First eliminate routes that cannot supply the duration, aspect ratio, resolution, audio and input controls your shots require. Then compare the remaining routes using the same model and settings where possible. A direct API and an aggregator can have different billing rules for a similarly named model. If you cannot establish equivalence of model version and settings, label the prices as distinct offers rather than claiming one is cheaper for identical output.

For fixed-length clips, budget the supported lengths shot by shot. For per-second routes, record the output duration actually billed. For token-priced Gemini video, inspect usage after a trial request and include any input and text-output tokens. For Luma, convert credits using your account rate. Record accepted results during a small sample and replace the 100%, 50% and 25% scenarios with your measured fraction. The AI model tracker can help you spot model-name changes; the provider’s current endpoint and price page remain the billing authority. Our text API price comparison explains why token input and output rates must stay separate, a distinction that also matters for Gemini Omni Flash.

Sources

• Runway Dev pricing — developer credit value and video-model rates.
• Google Cloud Model Garden pricing — Veo video and audio rates.
• Gemini API pricing — Omni Flash token rate and output-token conversion.
• BytePlus ModelArk pricing — published Seedance example and token-billing caveat.
• MiniMax pricing overview — Hailuo fixed clips, H3 rates, inputs and packages.
• Luma Agents pricing — Ray 3.2 credit table and request types.
• Higgsfield API — aggregator catalog and displayed Kling sale rate.
• fal video models and billing rules — example endpoint rate and charging rules.
• Replicate Veo 3.1 Fast — reseller endpoint and audio pricing.
• xAI API pricing — video output and input-image rates.

All pages read on November 6, 2026.

Free toolWhat will 10,000 chats a month cost?Estimate my bill