· reAPI Team
Video API pricing: input seconds vs output seconds
Understand video API pricing by separating input, output and compute seconds. Reconcile model rates, request settings and final charges before comparing costs.

Six cents per second looks cheaper than ten cents per second. For the same eight-second video, it can also produce the larger bill.
Take a hypothetical job with twelve seconds of reference footage and eight seconds of generated output. An output-only rate of ten cents gives a total of eighty cents. A six-cent rate applied to both input and output gives one dollar and twenty cents. These are invented rates for comparison, not prices offered by fal, Replicate or reAPI.
The arithmetic is simple once the unit is visible. The difficulty is that “per second” can mean several things in video API pricing. fal documents output-based, fixed-price and compute-based billing across its models; reAPI also documents reference-video jobs where input duration counts.[1][2] Before comparing the numbers on two model pages, establish which seconds each one multiplies.
Keep the source clip and the finished clip separate
In the hypothetical job, the twelve-second reference is material you supply. The eight-second output is material the model produces. Record both even if the model you currently use bills only one. Otherwise a later move to a different billing rule can make your estimate look mysteriously wrong.
Here is the same request under three illustrative offers:
| Illustrative offer | Quantity billed | Calculation | Total |
|---|---|---|---|
| Ten cents per output second | Eight output seconds | Eight × ten cents | Eighty cents |
| Six cents per input and output second | Twelve input + eight output seconds | Twenty × six cents | One dollar and twenty cents |
| Ninety cents for one eligible clip | One clip | One × ninety cents | Ninety cents |
All rates and eligibility assumptions in this table are hypothetical. The table compares arithmetic, not providers or output quality.
Now shorten the reference to four seconds while keeping the requested output at eight seconds. Under the output-only assumption, the total remains eighty cents. Under the input-plus-output assumption, it becomes seventy-two cents: twelve billed seconds at six cents each. The fixed-clip offer remains ninety cents if the modified request is still eligible for that offer.
This is why a recommendation to “use a shorter reference to save money” needs a condition. It changes the quantity in one calculation but not the others. It also changes what you ask the model to work from. The arithmetic does not establish that a shorter reference will preserve the motion or other information needed for your result.
Real models can add minimum windows or rounding rules. reAPI's task documentation explicitly notes model-dependent reference-video billing and minimum windows.[2] Read the selected model's rule before applying the simple sums above. If you cannot establish how a partial billing block is counted, your estimate still has an unanswered question; do not quietly choose the rounding that makes it cheaper.
For a specific reference-video implementation, the Seedance cost-per-second guide and model page provide the relevant context. Their rules should stay attached to that model and mode, rather than becoming a formula for every endpoint with “video” in its name.
Video analysis may bill footage without generating any
An analysis endpoint makes the distinction easier to see. You upload footage and receive text describing it. There is no output video whose duration could explain the charge.
The fal Video Understanding page retrieved on November 15, 2026 listed a price of $0.01 per five seconds. In its archived response to a billing complaint, fal explained that the counted duration was the input video's length, not the time spent generating the answer.[3] That reply explains the provider's interpretation of the unit. It does not independently verify the customer's disputed total.
Suppose, as a hypothetical timing example, a service analyzes a minute of footage in less than a minute. Under an input-duration rule, finishing sooner does not shorten the source clip. Conversely, under a compute-time rule, knowing that the source lasts a minute does not tell you how long the provider's measured processing interval lasted.
Keep the operation name beside the rate. Video understanding, text-to-video generation and reference-video generation are different jobs. A rate that looks low for analyzing input cannot be carried across to a model that produces new footage. The answer format and billed quantity both need to match the task you are budgeting.
Also retain the original unit when you normalize a quote. “One cent per five seconds of input” is more informative than a derived number labelled only “per second.” The original wording lets someone else check the conversion and identify whether a minimum block or other condition was left out.
Waiting time is another number again
Your application can measure the interval from submission to receiving a file. That interval is useful for user experience, but it is not automatically the billing interval. Queue waiting, active processing and file transfer occur at different stages.
fal states that time waiting in its Model API queue is not charged.[1] For time-based public models, Replicate describes billing for active processing. Most private models and deployments can instead be billed while their instances are setting up, active or idle, with an exception for fast-booting fine-tunes.[4] Check the execution type before using one rule to explain another service's total.
Return to the eight-second output. Even if your screen spends much longer waiting, you cannot substitute that elapsed wait for eight output seconds in the first row of the table. And if the service bills compute time, you cannot substitute eight output seconds for the measured processing interval. The word “seconds” is shared; the thing being measured is not.
Keep a timing record if you need to investigate latency. Keep the provider's usage record if you need to reconcile a charge. Joining them by the request ID is useful; treating them as the same measurement is what causes confusion. Dedicated capacity warrants a separate calculation, covered in the idle deployment cost guide.
Turn the quote into a request you can compare
The next comparison should describe one intended result. Write down the exact model and mode, source durations, output duration, resolution and audio setting. When a setting is omitted, look up its documented default. Do not assume two endpoints select the same default simply because your application sends the same short payload to both.
Then apply each offer's actual unit to that request. Express the quantity in the quoted unit: divide seconds by sixty for a per-minute rate, or use the documented block size and rounding rule for block billing. Keep any fixed charge separate. Only after that calculation should you compare the totals.
If the cheaper offer cannot accept the required reference footage, it has not priced the same job. If it produces a different output format or duration, decide whether that result still meets the brief. The first table intentionally leaves capability and quality unranked; a real provider choice cannot.
Save the estimate with its source date and assumptions. That gives you something specific to compare with the finished task. A bare “expected cost” cell is difficult to investigate later if nobody can tell whether it included input footage or which resolution it assumed.
fal provides pricing lookup, cost estimates, usage line items and analytics.[1] Use the available billing detail when investigating a fal request. A confusing quote is not evidence that the platform has no way to explain it.
Compare the estimate with settlement, not just the hold
On reAPI, a media submission reserves credits based on the parameters sent. Once the task settles, the final amount replaces that hold: excess reservation can return, while an amount above the reservation can be deducted from the remaining balance. The relationship depends on the model's billing rule.[2]
For ordinary media tasks, read the state alongside usage.credits. A processing task with zero settled usage has not finished billing. A completed task reports the amount charged. A failed task can show zero after a full refund or retain a positive amount under a model-specific post-generation policy.[2] Treating each of those zeroes as “free” would mix different events.
If the final amount differs from the estimate, compare the saved configuration and billed quantity first. Was the estimate output-only while the model counts reference footage? Did the model apply a documented minimum? Was a requested setting left at a different default? These are questions to resolve against the actual request and rule, not assumed explanations for every mismatch.
For support, supply the request ID, endpoint, time, relevant settings, your expected calculation and the final charge record. Show the particular quantity or rule that appears different. You do not need to share an API key or copy private footage into a billing spreadsheet to state that discrepancy.
Budget the clip you keep
The normalized total prices one attempt. A project budget also needs to count how many attempts produced the clip you ultimately use. Keep deliberate creative retakes separate from technical retries and from jobs whose results have not yet been recovered.
For another hypothetical calculation, use the output-only offer from the table: eighty cents per attempt. If three attempts are billed and one clip is accepted, the generation spend for that accepted clip is two dollars and forty cents. That is arithmetic with assumed outcomes, not a measured success rate. Replace the number of attempts with your own recorded work before using it in a forecast.
Do not assume every attempt without a usable clip is refundable. fal's FAQ distinguishes non-billable server errors from invalid-input errors that may be charged after GPU work. reAPI also documents model-specific post-generation safety failures that can retain charges.[5][2] The failed-request billing guide explains how to check those outcomes individually.
Delivery and storage belong in the project budget only where you actually incur them. They should not be invented as extra line items on a model's invoice. For recurring production, the daily video budget guide extends the calculation across repeated work.
For the next comparison, keep one row per attempt: model and mode, input duration, output duration, billing rule, final cost and whether the clip was accepted. Then the eight-second clip on your screen has a traceable cost behind it, whichever number the pricing page places beside “per second.”
References
- fal. Model API Pricing: Billing Units and Pricing API. Retrieved November 15, 2026.
- reAPI. Tasks: usage, reservation, status and refund behavior. Read November 15, 2026 from reapi.ai/docs/api/tasks.
- fal. Video Understanding endpoint pricing (retrieved November 15, 2026); fal business response to a video-understanding billing review (April 10, 2026, archived November 15, 2026).
- Replicate. Billing: public models, private models, deployments and instance lifecycle. Retrieved November 15, 2026.
- fal. Model APIs FAQ: failed requests and billing. Retrieved November 15, 2026.