// unified video generation api · early access

One endpoint for every video model

Veo, Kling, Runway, Seedance, Wan, Hailuo and the rest behind a single schema. Below: the full comparison table, real per-second pricing, latency numbers and the async pattern every video API shares.

12 models, one keyper-second billingwebhooks + polling
01

The model catalog

Every generative video model that matters in July 2026, with its list price band. Click nothing; the full spec table is one scroll below.

Google
Veo 3.1
$0.05 – 0.20 /s
audioi2v1080p
Kuaishou
Kling 3.0 Pro
$0.084 – 0.112 /s
audioi2v15s clips
Runway
Gen-4.5
$0.20 – 0.25 /s
silenti2vquality #1
ByteDance
Seedance 2.0
$0.08 – 0.10 /s
audio9 ref imgs
Alibaba
Wan 2.7
$0.02 – 0.10 /s
silentopen weightscheapest 1080p
MiniMax
Hailuo 2.3 Pro
~$0.05 – 0.08 /s
silenti2vvalue
Luma
Ray 3
~$0.06 – 0.12 /s
silenti2vextend
Amazon
Nova Reel
~$0.08 /s
720p2min multi-shot
PixVerse
v6
~$0.04 – 0.08 /s
silenti2v
Vidu
Q3
~$0.04 – 0.08 /s
16s @ 1080p
Tencent
Hunyuan Video
compute only
open sourceself-host
OpenAI
Sora 2
credit-based
API retiring 09/2026audio
02

Full comparison: price, latency, specs

List pricing and latency compiled from provider docs and public aggregator rates, checked July 2026. Prices are per second of generated video; the 10s-clip column keeps the math honest.

Model$/second10s clipLatencyMax resMax lengthAudioImage-to-videoAccess
Google Veo 3.1$0.05 to 0.20$0.50 to 2.0015 to 25s1080p8s (+extend)YesYesGemini API, Vertex, aggregators
Kling 3.0 Pro$0.08 to 0.11$0.84 to 1.1215 to 30s1080p15sYes (Omni)YesKling API, aggregators
Runway Gen-4.5$0.20 to 0.25$2.00 to 2.5020 to 40s1080p10s (+extend)NoYesRunway API, aggregators
Seedance 2.0$0.08 to 0.10$0.80 to 1.00~45s+1080p10sYesYes, 9 refsBytePlus, aggregators
Wan 2.7$0.02 to 0.10$0.20 to 1.0030 to 60s1080p10sNoYesAlibaba, aggregators, self-host
Hailuo 2.3 Pro~$0.05 to 0.08~$0.50 to 0.8030 to 60s1080p10sNoYesMiniMax API, aggregators
Luma Ray 3~$0.06 to 0.12~$0.60 to 1.2020 to 40s1080p10s (+extend)NoYesLuma API, aggregators
Amazon Nova Reel~$0.08~$0.8060s+720p2min multi-shotNoYesAWS Bedrock
PixVerse v6~$0.04 to 0.08~$0.40 to 0.8020 to 40s1080p8sNoYesPixVerse API, aggregators
Vidu Q3~$0.04 to 0.08~$0.40 to 0.8030 to 60s1080p16sNoYesVidu API, aggregators
Hunyuan VideoCompute onlyGPU costHardware-bound720p+~5sNoYesSelf-host, open weights
OpenAI Sora 2Credit-basedVaries60s+1080p20s (+ext 120s)YesYesOpenAI API (deprecated)
latency = typical time to first downloadable clip for a 5-10s generation · ranges reflect resolution tiers and direct-vs-aggregator differences
deprecation notice · sora 2 videos api

OpenAI has announced the Sora 2 Videos API shuts down on September 24, 2026. If you built on it, the practical migrations are Veo 3.1 (closest on quality plus native audio), Kling 3.0 (cheaper, faster) or an aggregation layer so the next deprecation is a config change instead of a rewrite. That last option is exactly why this unified API exists.

03

Three kinds of "video API"

Half the frustration in this market is vendors using one term for three different products. Route yourself first.

generative

Text/image to video models

Veo, Kling, Runway, Seedance, Wan, Hailuo. A prompt or a still image goes in, novel footage comes out. Priced per second of output. This page (and this API) is about these.

avatar

Script to presenter

HeyGen, Synthesia, D-ID, Tavus. A script becomes a talking-head video with a licensed or cloned presenter. Priced per minute or credit. Right choice for training and personalized outreach video.

render

Template assembly

Shotstack, Creatomate, JSON2Video. Programmatic editing of your existing clips, images and captions. Nothing is generated. Right choice for automated slideshows and data-driven video.

04

The async pattern every video API shares

Video generation takes 15 to 60+ seconds, so every serious API is asynchronous: submit a job, then poll or receive a webhook. The same integration in three languages; swap the model string to switch providers, nothing else changes.

curl
# 1. submit
curl -X POST .../v1/generate \
 -H "Authorization: Bearer $KEY" \
 -d '{"model":"kling-3.0",
     "prompt":"product on marble
      pedestal, studio light",
     "duration":5}'

# 2. poll until done
curl .../v1/jobs/gen_8f2k1 \
 -H "Authorization: Bearer $KEY"
# {"status":"succeeded",
#  "video_url":"https://..."}
python
import time, requests

job = requests.post(
  f"{BASE}/v1/generate",
  headers=auth,
  json={"model": "veo-3.1",
        "prompt": prompt,
        "duration": 8}
).json()

while True:
    j = requests.get(
      f"{BASE}/v1/jobs/{job['id']}",
      headers=auth).json()
    if j["status"] != "running":
        break
    time.sleep(3)

print(j["video_url"])
node
const job = await fetch(
  `${BASE}/v1/generate`, {
  method: "POST",
  headers: auth,
  body: JSON.stringify({
    model: "wan-2.7",
    prompt,
    duration: 5,
    webhook_url:
      "https://app.dev/hooks"
  })
}).then(r => r.json());

// webhook fires on completion:
// { id, status: "succeeded",
//   video_url, seconds_billed }

Two production notes the docs never lead with. First, implement the webhook path even if you start with polling: long generations plus polling loops are how serverless bills explode. Second, store the provider job ID before you await anything; when a function times out mid-poll, that ID is the difference between resuming and re-billing.

05

What 1,000 ten-second clips actually cost

Per-second prices look interchangeable until you multiply. The same monthly workload priced across the field.

Model1,000 x 10s clips / monthRead
Wan 2.7$200 to 1,000The volume play. B-roll, drafts, high-iteration creative testing.
Hailuo 2.3 / PixVerse / Vidu$400 to 800The value middle. Good quality per dollar for social content.
Seedance 2.0$800 to 1,000Pay for multimodal control: 9 reference images, audio input.
Kling 3.0 Pro$840 to 1,120Fastest latency tier plus audio. The ads workhorse.
Veo 3.1$500 to 2,000Cinematic ceiling; the range is the resolution tier you pick.
Runway Gen-4.5$2,000 to 2,500Benchmark-topping quality. Hero shots, not volume.

The strategy that falls out of this table: route by shot value. Draft and iterate on a cheap model, re-render the winning prompt on a premium one. Teams doing this cut video spend 60 to 80 percent versus running everything on the flagship, and it is precisely the workflow a unified API makes a one-line change.

06

Which model for which job

You're buildingUseWhy
Ad creative testing at volumeKling 3.0, Seedance 2.0Cheap enough to test 20 variants, fast enough to iterate same-day, native audio for sound-on placements.
Cinematic brand filmVeo 3.1, Runway Gen-4.5Best physics, coherence and camera language. Render few, render high.
Social b-roll pipelinesWan 2.7, Hailuo 2.3Cost per clip low enough to generate daily at feed scale.
Product demos from stillsKling 3.0, PixVerse v6Strong image-to-video: real product photos become motion without a studio.
Avatar or training videoHeyGen, Synthesia, TavusDifferent category (section 03). Script-to-presenter beats generative models for talking heads.
Full control / no per-second feesHunyuan, Wan (open weights)Self-host on your GPUs. You trade ops burden for marginal cost.

Direct integration vs an aggregator layer

Going direct to one provider gets you enterprise SLAs, the newest checkpoints first, and one vendor relationship. It also gets you their rate limits, their deprecation schedule (see Sora) and a rewrite every time the leaderboard flips, which in this market is roughly quarterly.

An aggregation layer (one key, one schema, model as a string parameter) trades a small markup for portability: reroute when prices drop, A/B models on real traffic, and survive deprecations with a config change. Our position is obviously the second camp, since that is what we are building, but the honest rule is: direct if you are certain of your model and volume-negotiating with the vendor; aggregate if you value optionality. Most teams shipping product, not research, want optionality.

07

Video generation API questions, answered

Q.01What is the cheapest AI video generation API?+

Wan 2.7 is the cheapest mainstream 1080p option at roughly $0.02 to 0.10 per second depending on host and tier, which puts a 10-second clip at $0.20 to 1.00. If you can run your own GPUs, open-weight models (Wan, Hunyuan) drop marginal cost to pure compute. For hosted volume work, the Hailuo and PixVerse tier around $0.04 to 0.08 per second is the usual sweet spot.

Q.02What is the fastest video generation API?+

Veo 3.1 and Kling 3.0 currently lead on latency, typically returning a 5 to 10 second clip in 15 to 30 seconds. Seedance and Wan trade speed for cost or control, usually landing at 30 to 60+ seconds. Latency varies with load, resolution and duration; benchmark with your own prompts at your usual hours before committing.

Q.03Which video APIs generate audio too?+

Veo 3.1, Kling 3.0 (Omni), Seedance 2.0 and Sora 2 generate synchronized audio natively, including ambient sound and speech. Runway, Wan, Hailuo and Luma output silent video; you add sound in post or via a TTS pass. For sound-on placements like TikTok, native audio saves a pipeline step and usually sounds more coherent.

Q.04Is the Sora API still available?+

OpenAI has announced the Sora 2 Videos API shuts down on September 24, 2026. Existing integrations keep working until then. If you are choosing today, build against Veo 3.1 or Kling 3.0, or integrate through an abstraction layer so the next provider change is configuration, not code.

Q.05How is video API pricing billed: per second or credits?+

Most providers bill per second of generated output, with the rate depending on model and resolution tier. Some (Runway, Sora) use credit systems that abstract the same thing. Aggregators normalize everything to per-second or per-clip pricing. Watch for the failed-generation policy: good APIs don't bill failed or moderation-blocked jobs, but not all are good.

Q.06Can I use generated videos commercially?+

Hosted providers (Google, Kling, Runway, MiniMax) grant commercial usage rights on paid tiers, and none of the majors watermark paid API output. Open-weight models vary: Wan's license permits commercial use, others restrict above certain scale. Whatever you ship, keep the generation metadata; provenance requirements (C2PA labeling) are tightening across ad platforms.

Q.07What resolution and length can these APIs produce?+

The current hosted standard is 1080p at 5 to 15 seconds per generation, with extension endpoints chaining clips toward 30 to 120 seconds. Nova Reel does multi-shot two-minute videos at 720p. True 4K generation is not commercially standard yet; teams upscale 1080p output when they need it.

Q.08Can I self-host a video generation model?+

Yes: Hunyuan Video and Wan publish open weights that run on a single high-memory GPU (think H100 or a rented A100 class card). You give up managed scaling, safety filtering and the newest checkpoints, and gain unlimited generation at fixed hardware cost. Break-even versus API pricing typically lands around several thousand clips per month.

Q.09Should I integrate one provider directly or use an aggregator?+

Direct if you have negotiated volume pricing with one vendor and stability matters more than flexibility. Aggregator if you want to route between models, hedge deprecations and test quality-per-dollar continuously. The market leaderboard has flipped roughly every quarter since 2024, which is the strongest argument for keeping the model name a string in your config.

Q.10What is the best AI video generation API overall?+

Runway Gen-4.5 tops quality benchmarks, Veo 3.1 leads the cinematic-plus-audio combination, Kling 3.0 wins price-to-performance, and Wan 2.7 wins pure price. There is no single best, which is the point of this page: match the table above to your use case, or use a unified API and stop betting your roadmap on one vendor.

One key. Every video model.

We're onboarding early-access developers now. Unified schema, per-second billing, webhooks, and model routing across everything in the table above.