Video generation is the youngest category in our catalog and the thinnest: exactly two routes exist today. Kling, hosted via fal.ai, and Wan 2.2, hosted via Replicate. When a category has two options, a comparison post writes itself, but this one is worth doing carefully because video pricing works differently from every other category we route, and the operational realities matter as much as the numbers.
The numbers
Video is priced per clip-second: a ten-second clip costs ten times a one-second clip. Routed prices (provider list plus 20%, per /docs/pricing):
| Model / host | Price per clip-second | Quality score |
|---|---|---|
| Kling via fal.ai | $0.048 | 84 |
| Wan 2.2 via Replicate | $0.06 | 80 |
Kling is both cheaper and higher-scored in our catalog, which makes the headline verdict unusually clean: Kling is the default, Wan is the alternative. Most comparisons I write end in “it depends.” This one mostly does not, and I will spend the rest of the post explaining the qualifications hiding inside “mostly.”
Per-second math at real clip lengths
Per-second prices are small enough to feel free and multiply fast enough not to be. At typical clip lengths:
| Clip length | Kling (routed) | Wan 2.2 (routed) |
|---|---|---|
| 5 seconds | $0.24 | $0.30 |
| 10 seconds | $0.48 | $0.60 |
Now put that next to image generation, where the floor in our catalog is about a penny per image: one 10-second clip costs as much as roughly 44 cheap images. If your agent generates video speculatively (the way agents happily fire search queries), the bill will teach you a lesson. Video is the category where “generate three variants and pick the best” triples a meaningful number, and where prompt quality upstream pays for itself immediately. Batch-of-one, review, iterate.
The 25% gap between the two providers compounds the same way. At 1,000 ten-second clips a month, Kling costs $480 routed and Wan costs $600. Real money, but notice both numbers are small next to the human time spent reviewing a thousand clips. Video costs are rarely the binding constraint; attention is.
The async job reality
Every other category I route feels like a function call: request in, result out, milliseconds to seconds. Video does not work like that. Generation takes long enough that these are jobs, not calls, and your integration has to be built around that fact.
Our router gives video the longest per-request wall-clock budget of any category: 330 seconds, versus 20 for search. That number should recalibrate your architecture. An agent cannot block a conversational turn on a video render; you kick off generation, go do something else, and collect the result. Plan for queue-shaped code (polling or webhooks, retries that respect money already spent, idempotency so a retry does not render twice), because a failed video job at second 300 is a different beast from a failed search query at millisecond 200.
This is also where failover semantics get interesting. With two providers in the category, a Kling failure fails over to Wan, and the router tries up to three providers within that 330-second budget. But video failover means re-rendering from scratch on a different model, so the output will differ, not just the letterhead. For agents that need a video, any video, automatic failover is exactly right. For workflows where the model’s look is the point, you may want to pin one provider via saved routing preferences and handle failures by waiting instead.
Which model when
The honest quality caveat: our scores (84 and 80) are hand-curated ratings, and video quality is the most prompt-dependent and taste-dependent judgment in the catalog. The four-point gap says Kling output rates better in our evaluation overall; it does not say Kling wins on your prompt. The two models have visibly different characters, and for a specific aesthetic, in my experience, it is entirely possible to prefer the lower-scored one.
So the practical guidance:
- Default and cost-sensitive traffic: Kling via fal. Cheaper and higher-scored is not a hard call.
- Failover and diversity: Wan via Replicate. A second independent host-and-model pair is worth having wired up before you need it.
- Style-driven work: test both on your actual prompts before pinning. Two providers is a small enough field that an afternoon of side-by-side renders settles it empirically.
Worth knowing: the category is still filling in (Modal is catalogued as coming soon on the sandbox side, and I expect the video roster to grow the same way), so decisions here are worth revisiting quarterly. I keep broader category notes in the best video generation APIs roundup.
Both routes, priced per clip-second from the same open catalog our router reads, are side by side on the video generation comparison page.