Video generation is live on Kyma
Kyma now generates video. The same API key that hits chat/completions also hits videos/generations. Five models, per-second billing, one ledger, and a single async endpoint that flips between text-to-video and image-to-video by passing an image_url.
Every clip in this post was generated on Kyma. The hero, the multi-shot action, the long-form b-roll, the product ad - all dogfood, no stock.
Kling 3 Pro Audio, 5 seconds, $1.1340. Native dialogue and ambient audio rendered in one pass with the visuals - no muxing, no separate sound-design step.
What's available
| Model | Best for | Per second |
|---|---|---|
| kling-3-pro | Cinematic hero, character shots, premium brand video | $0.1512 |
| kling-3-pro-audio | Cinematic clips with native ambient or dialogue | $0.2268 |
| kling-2.5-pro | Budget cinematic, b-roll, character shots (5 or 10s) | $0.0945 |
| seedance-2-pro | Multi-shot action, social with synced audio, product motion | $0.410 |
| seedance-2-fast | Social shorts, UI motion, rapid iteration | $0.327 |
Pricing is per second of generated footage, billed against the actual clamped duration. A 5-second Kling 3 Pro clip is $0.7560. A 5-second Seedance 2 Fast is $1.6350. No subscription, no per-frame surcharge.
Showcase
Kling 3 Pro Audio - native dialogue, no muxing
The hero clip above is the model's actual differentiator. Voice, ambient sound, and visuals - rendered together, synchronized, in one pass. Most pipelines need a separate sound-design step. With kling-3-pro-audio, you skip it.
The premium over kling-3-pro is exactly 50% per second. If you need diegetic sound, take the audio variant. If you don't, the silent one is cheaper.
Seedance 2 Pro - multi-shot action
Seedance 2 Pro, 5 seconds, $2.0480. Prompt: multi-shot urban skater sequence - first a low tracking shot of wheels rolling on wet asphalt at dusk, then cut to a wide shot of the rider carving through neon-lit alleys.
Seedance 2 Pro is the model to reach for when a single static shot isn't enough. It interprets multi-shot prompts and renders the cuts in-frame, with native audio bundled. 720p output, dynamic camera moves, and the same image_url parameter for image-to-video if you want to lock the opening frame.
Seedance 2 Fast - long-form b-roll
Seedance 2 Fast, 15 seconds, $4.9050. Underwater dragon dive, full duration on the cheaper tier. The same SKU stretches from 5-second iteration to 15-second hero.
Seedance 2 Fast - product ads
Seedance 2 Fast, 5 seconds, $1.6350. Real Kimchi product ad - jar hero, motion typography, native audio. The same family as Pro at ~20% lower cost per second. Good for UI motion, product demos, social shorts where you'd rather iterate three takes than wait for one.
Why these five (and not twenty)
We picked five models that don't overlap. Each one does a thing the others can't:
- •Kling 3 Pro is the photoreal frontier for video. Smooth motion, sharp humans, the kind of clip you'd put above the fold.
- •Kling 3 Pro Audio is Kling 3 Pro plus synchronized sound, in one pass. Skip the post-production muxing.
- •Kling 2.5 Pro is the budget cinematic sibling. Locked at 5 or 10 seconds, but at $0.0945 per second it's the cheapest way to get a Kling-quality b-roll clip.
- •Seedance 2 Pro does multi-shot action with native audio. Different shape of output - if Kling renders one beautiful shot, Seedance renders a sequence.
- •Seedance 2 Fast is the iteration tier. ~20% cheaper, faster, same family behavior. Use it when you'd rather try ten variations than wait on one Pro render.
Adding a sixth photoreal model wouldn't have given anyone a real choice. We'd rather route five sharp tools than twenty fuzzy ones.
How it works
Same OpenAI-compatible base URL. Same key. POST to /v1/videos/generations, get a job_id back, poll GET /v1/jobs/{id} until succeeded. Most clips finish in 1-3 minutes.
A few behaviors worth knowing:
- •It's async. Submit returns a
job_idand a 202 immediately. The runner times out at 10 minutes. - •Pricing is held at submit, charged at success. Kyma places a hold of
per_second × durationat submit time. If the upstream fails, the hold is released and you never pay for a failed generation. - •Output URLs are stable. We rehost the result on our own blob storage so the URL doesn't expire when the upstream cleans up its cache. The URL in the API response is the one you can embed.
- •Idempotency keys are first-class. Replay the same
(api_key, idempotency_key)pair and you get the same job back, no duplicate charge. - •T2V or I2V from the same call. Pass
image_urlin the body and the request switches to image-to-video — your image becomes the first frame and the prompt drives the motion. All five models accept it. Pairs naturally withflux-1.1-ultrato render a hero frame, then animate it withseedance-2-proorkling-3-pro. One API key, one billing ledger.
Full request shape, parameters, polling, and SDK examples are in the video generation API reference.
Duration limits and pricing
Per-SKU duration caps:
| Model | Min | Max | Notes |
|---|---|---|---|
| kling-2.5-pro | 5 | 10 | Snap-to-5 or 10 only |
| kling-3-pro | 3 | 15 | Any integer |
| kling-3-pro-audio | 3 | 15 | Any integer |
| seedance-2-pro | 4 | 15 | Any integer |
| seedance-2-fast | 4 | 15 | Any integer |
A 15-second Seedance 2 Pro clip is $6.15. A 15-second Kling 3 Pro is $2.27. A 5-second Kling 2.5 Pro b-roll clip is $0.47. Pick the SKU and the duration that match what the asset is for.
Use cases this unlocks
- •Marketing automation. Per-product hero clips on demand instead of a video shoot per asset.
- •Social pipelines. Seedance 2 Fast for the daily TikTok or Reels, Seedance 2 Pro for the weekly campaign.
- •Product motion. UI explainers, onboarding loops, error-state animations - rendered fresh, on-brand, with the same key your app already has.
- •Image-to-video animation. Render a still with
flux-1.1-ultra, animate it withkling-3-proorseedance-2-pro. Your storyboard becomes a sequence in one pipeline. - •Talking-head and ambient sound.
kling-3-pro-audiofor content that needs diegetic audio without a separate sound-design pass.
What's next
We're watching production behavior over the next week - tail latency on long clips, refund rate, blob URL stability, the shape of T2V vs I2V traffic. The model pages at /models/video-generation stay current as we tune the routing.
Phase 3 ideas on the roadmap: longer durations as upstream caps grow, audio-only generation, lipsync from a still + script. We'll publish dates when we have provider parity verified.
For now, this is what's live. Same key, five models, per-second pricing, hold-and-finalize billing, T2V or I2V from one endpoint.
Links
- •Video generation API reference
- •Video models page
- •Image generation blog
- •Pricing
- •Sign up - free signup credit covers a few short clips