Kling 4.0 API: Price per Second, Parameters, Python Example

Kling 4.0 is available as an API on Orux AI under the model name kling-4.0, priced at 42.75 Credits ($0.214) per second of video at 720p and 63.75 Credits ($0.319) per second at 1080p. A 5-second 720p clip costs 213.75 Credits ($1.07), and you only pay when the task succeeds. Third-party trackers report lower per-second prices for Kling 4.0 elsewhere, and this article shows those numbers next to ours, along with every request parameter, a Python script that survives restarts, and two real example clips.

Is there a Kling 4.0 API?

Yes. Kuaishou released Kling 4.0 in October 2026, and it has been listed on Orux AI since late October as kling-4.0. On Orux it runs through the same asynchronous task API as every other video and image model: you POST a task, poll its status, and download the result. One key, one balance, and the same code you would use for Seedance or Wan.

Kling 4.0 accepts text plus optional images (first and last frames, keyframes, reference images and named subjects) and returns video. It supports durations from 3 to 30 seconds at 720p or 1080p. Video input, such as video references or editing an existing clip, is not available on Orux AI yet, and neither is automatic duration: you choose an exact length.

Kling 4.0 price per second

Orux AI bills Kling 4.0 by output second, with the rate set by resolution. Nothing else (aspect ratio, reference images, subjects or voices) changes the price. Prices as of Oct 29, 2026; 1 Credit = $0.005, dollar values for comparison only.

Swipe horizontally to view the full table.

Now the comparison you should see before you buy. At the time of writing, third-party price trackers report Kling 4.0 on fal at about $0.157 per second at 720p and $0.235 per second at 1080p, with fal defaulting to 1080p (AIReiter, updated Oct 24, 2026; LLMReference, refreshed Oct 21, 2026). We could not open fal's own Kling 4.0 page to confirm, so treat those as reported numbers.

Swipe horizontally to view the full table.

If those reported prices hold, Orux AI is about 36% more expensive than fal for Kling 4.0 at both resolutions. We are not going to call it cheap. Kuaishou's own API price for Kling 4.0 was not available for us to verify at the time of writing, so we do not list one.

What you get on Orux AI for the difference: a single key and task API covering Kling, Seedance, Wan and Vidu video models plus image and chat models; no charge for failed or expired tasks; and one prepaid balance in Credits with no monthly fee. If you only ever call Kling 4.0, the reported fal price is lower. If you run several video models side by side, that is the trade-off to weigh.

If Kling 4.0 is more than you need, the older Kling models on Orux AI cost much less per second:

Swipe horizontally to view the full table.

For the full per-second ranking across video models, see cheapest AI video generation API, and for ByteDance's line-up, Seedance 2.0 API pricing.

Request parameters

Everything goes inside the task's input object.

Swipe horizontally to view the full table.

A few practical notes. First and last frames are the most reliable way to control where a shot starts and ends, which matters if you are cutting several clips together. Named subjects (elements) are for keeping a character or product consistent across shots; give each one 2 to 4 clear images from different angles. All image inputs must be publicly reachable URLs, not local paths.

Python: submit, poll, download

Video tasks take minutes, not seconds, so the script polls every 10 seconds. It also writes the task ID to disk before polling. If the process crashes or you redeploy mid-render, rerunning the script resumes the same task instead of submitting a second paid one.

# pip install requests
import json, os, pathlib, time, requests

BASE = "https://orux.top/api/v1"
HEADERS = {"Authorization": f"Bearer {os.environ['ORUX_API_KEY']}"}
STATE = pathlib.Path("kling_task.json")  # survives restarts, prevents double submission

RATE = {"720p": 42.75, "1080p": 63.75}  # kling-4.0, Credits per second, checked 2026-10-29

job = {
    "model": "kling-4.0",
    "input": {
        "prompt": "A slow cinematic camera push towards a ceramic mug on a wooden desk "
                  "in soft morning light. No people, no text.",
        "resolution": "720p",
        "duration": 5,
        "aspect_ratio": "16:9",
    },
}
cost = RATE[job["input"]["resolution"]] * job["input"]["duration"]
print(f"estimated charge on success: {cost} Credits (${cost * 0.005:.3f})")  # 213.75 Credits

if STATE.exists():
    task_id = json.loads(STATE.read_text())["task_id"]
    print("resuming", task_id)
else:
    r = requests.post(f"{BASE}/tasks", headers=HEADERS, json=job, timeout=60)
    if r.status_code == 402:
        raise SystemExit("Balance too low: top up before submitting")
    r.raise_for_status()  # expect 202 Accepted; 429 means back off and retry later
    task_id = r.json()["task_id"]
    STATE.write_text(json.dumps({"task_id": task_id}))
    print("submitted", task_id)

while True:
    t = requests.get(f"{BASE}/tasks/{task_id}", headers=HEADERS, timeout=30).json()
    if t["status"] not in ("QUEUED", "RUNNING"):
        break
    time.sleep(10)

if t["status"] != "SUCCESS":
    STATE.unlink()  # FAILED / EXPIRED are not charged; safe to resubmit
    raise RuntimeError(t.get("error") or t["status"])

print("result link expires in:", t.get("result_expires_in"))
for i, url in enumerate(t["result"]):
    with requests.get(url, stream=True, timeout=300) as resp:
        resp.raise_for_status()
        with open(f"kling_{task_id}_{i}.mp4", "wb") as f:
            for chunk in resp.iter_content(chunk_size=1 << 20):
                f.write(chunk)
STATE.unlink()
print("saved")

Two things this script does on purpose. It deletes the state file only after a successful download or a terminal failure, so a crash during download does not lose a paid result. And it downloads immediately, because result URLs are temporary; there is no callback on Orux AI, so polling is the only way to know a task is done.

To start from a fixed first frame, add "first_frame_url": "https://example.com/desk.jpg" to input and mention @first_frame in the prompt.

Same prompt, two different shots

Before you script a hundred renders, look at what one prompt actually produces. These are two existing Kling 4.0 examples on the model page, not new generations for this article. Both store the same prompt and settings: 3 seconds, 720p, 16:9, no reference image. Both files are 1280 × 720, about 3.04 seconds long, and each contains an audio track.

The prompt:

A slow cinematic camera push towards a ceramic mug on a wooden desk in soft morning light. No people, no text.

At about the one-second mark, A shows a taller mug with a narrower opening, and B a rounder mug with different rim and handle proportions. The wood grain and light differ too. "A ceramic mug" names a category, not a product, so the model is free to invent one. The example names are labels, not a controlled test of keyframe versus reference modes. For real product work, pass the product through image_urls or elements and check its shape and markings before anything else.

To choose between two takes, we use three checks:

• Composition fits the job: decide where the headline or logo goes, then check that space in the final crop, not just in the full landscape frame.
• Motion is continuous: play the whole clip and watch the rim, handle and wood texture through the push. A still frame cannot show temporal stability.
• The ending connects: if an end card or another shot follows, check the last frames for a clean hold or cut, not just the best-looking opening frame.

When iterating, keep duration and aspect ratio fixed and change one instruction at a time, camera first, lighting in a separate round. Two examples cannot tell you a success rate; they only show that identical inputs can give visibly different shots.

Cost control tips

• Draft at 720p and 3 to 5 seconds. A 3-second 720p test costs 128.25 Credits; the same test at 30 seconds in 1080p costs 1,912.5 Credits, nearly 15 times as much.
• Decide the subject with stills first. Settling product look or character design with an image model costs a fraction of a video render.
• Move to 1080p only for the final take. Resolution is the only setting besides duration that changes the price.
• Use Kling v3 or v2.6 for simple text-to-video. If you do not need frames, references or clips longer than 15 seconds, Kling v3 std at 20 Credits per second costs less than half as much.
• Persist task IDs. Accidental double submission is the most common avoidable cost in video pipelines.
• Failed and expired tasks are free, so retrying a failure costs nothing extra; resubmitting a success does.

FAQ

Q.How much does Kling 4.0 cost per second?

On Orux AI, Kling 4.0 costs 42.75 Credits ($0.214) per second at 720p and 63.75 Credits ($0.319) per second at 1080p, as of Oct 29, 2026. A 5-second 720p clip is 213.75 Credits ($1.07). Third-party trackers report fal at about $0.157 and $0.235 per second, roughly 36% lower.

Q.What durations and resolutions are supported?

Kling 4.0 on Orux AI accepts any whole-number duration from 3 to 30 seconds, at 720p or 1080p, with aspect ratios auto, 16:9, 9:16, 1:1 and 21:9. You must choose an exact duration; automatic duration and video references are not available yet.

Q.Does Kling 4.0 generate audio?

Both Kling 4.0 example files on Orux AI contain an audio track, and the request accepts up to 3 named voices. Pricing has no separate audio tier: the per-second rate depends only on resolution. That differs from Kling v3, where turning sound on raises the price from 20 to 30 Credits per second in std mode.

Q.Kling 4.0 vs Kling v3: which should I use?

Use Kling v3 for straightforward text-to-video up to 15 seconds, at 20 Credits per second in std mode. Use Kling 4.0 when you need first and last frames, keyframes, reference images, named subjects or clips longer than 15 seconds. We have not run a controlled quality comparison between them.

Q.Am I charged for failed videos?

No. Video tasks on Orux AI are billed only when they reach SUCCESS. Tasks that end as FAILED or EXPIRED cost nothing, so you can retry them. Save the task ID right after submitting so a crash does not lead you to pay for a second, duplicate render.

Sources and next step

Play both examples and load their settings on the Kling 4.0 model page; loading a template does not start a paid generation. Orux AI prices checked on 2026-10-29. Sources: AIReiter Kling 4.0 pricing, LLMReference Kling 4.0 on fal, Orux AI pricing, task API docs.