Video generation is the youngest category I route, and the most honest thing I can say about it upfront is that the list is short: our catalog currently routes two providers. That is not a curation flourish; it reflects a category where API-accessible, production-billable text-to-video is still thin on the ground compared to, say, image generation’s five hosts. Short lists deserve honest treatment, so here is what the two options actually are, what they cost per second, and what nobody’s landing page tells you about running video generation inside an agent.
How video APIs bill: the clip-second
First, the unit. Video APIs bill per second of generated clip, not per request. A 5-second clip at $0.06 per clip-second costs $0.30; a 10-second clip costs $0.60. This matters because clip length is usually a parameter you control, which makes video the rare category where your request shape directly sets your unit cost. Budgeting per clip requires deciding your clip length first. (Routed prices below are provider list plus 20%, per the pricing docs.)
1. Kling via fal.ai: $0.048 per clip-second
The quality pick of our pair, catalog score 84. Kling is one of the text-to-video models that made the leap from demo reel to usable API, and fal.ai hosts it the way fal hosts most media models: as a managed inference endpoint with the job-queue machinery handled for you. At $0.048 per clip-second routed, a 5-second clip runs $0.24 and a 10-second clip $0.48.
fal is also the stronger platform bet of the two hosts if your product already generates images, since it serves FLUX variants as well (covered in my FLUX host comparison), and consolidating media inference on one vendor has real operational value.
2. Wan-2.2 via Replicate: $0.06 per clip-second
Wan-2.2 is an open-weight video model, hosted here by Replicate at $0.06 per clip-second routed, catalog score 80. A 5-second clip costs $0.30. It is both the pricier option and the lower-scored one in our catalog, which is an unusual combination, and I will not pretend otherwise: on pure price-per-quality, Kling via fal wins this matchup today.
So why route Wan at all? Failover, mostly. In a two-provider category, the second provider is your entire resilience story, and video generations are long-running jobs where a provider incident strands work in a way a failed search never does. There is also the open-weights angle: Wan’s weights being open means its hosting market can widen and its price can move the way FLUX hosting did. Worth watching, not yet worth switching for.
The table
| Model / host | Routed price per clip-second | 5s clip | 10s clip | Catalog score |
|---|---|---|---|---|
| Kling via fal.ai | $0.048 | $0.24 | $0.48 | 84 |
| Wan-2.2 via Replicate | $0.06 | $0.30 | $0.60 | 80 |
The async reality
Here is the part that bites agent builders. Every other category I route returns in seconds; video does not. Generation is a genuinely long job, and our router’s per-request wall-clock budgets tell the story in one line: search gets 20 seconds, video gets 330. That is not padding, it is the honest ceiling for how long you may wait for a clip.
Designing an agent around a call that can run minutes is different work:
- Do not block the loop. An agent that synchronously awaits a video generation is frozen for the duration. Kick the job off, continue, collect the result.
- Failover is expensive here. When a search fails over, you lose a second. When a video job dies late and the router retries elsewhere, the retry costs minutes of wall clock. This is why the circuit breaker matters more in slow categories: skipping a provider that is above a 30% error rate over 5 minutes (the router’s threshold) saves you the single most expensive failed attempt in the catalog.
- Watch the multiplication. An agent that “just tries a few variations” at 4 clips per idea and $0.24 to $0.30 per 5-second clip is spending a dollar-plus per idea without anyone deciding to. Per-clip costs feel small; per-loop costs are what show up on the bill I pay.
Honest limits of the category
Clips are short, single-shot, and prompt-sensitive; nobody is generating a coherent minute of narrative video from one API call, and quality varies with prompt in ways our two catalog scores cannot fully capture. Treat 2026 hosted text-to-video as a b-roll and asset generator, not a film studio. That is still genuinely useful (thumbnails in motion, product shots, social clips), and it is priced like the early-market capability it is.
Both providers, current prices, and quality scores live on the video generation comparison page, generated from the same open catalog the router reads.