Seedance vs Kling vs Hailuo vs Wan: AI Video Models, One Prompt

Key takeaways
- Five AI video models each got the same prompt once, and we compared the raw outputs.
- Wan 3.0 built the most detailed setting; the two Seedance models followed the camera direction most closely.
- Kling 3.0 Pro was the only one to return 1080p, and Hailuo 2.3 the only one with no audio track.
- Seedance 2.0 Fast, Hailuo 2.3 and Wan 3.0 finished the latte art on screen; the other two ended mid-pour.
- One run per model is a snapshot: the same model gives a different clip on every run.
Each video model has its own idea of what a shot should look like. We ran five of the models available in FastMetal Studio and the API on the same prompt on November 17, 2026.
The test
One prompt, sent in English:
A barista in a small Tokyo coffee shop pours latte art into a white cup, steam rising, soft morning light through the window, the camera slowly pushes in toward the cup. Realistic, cinematic.
It packs in hand movement, liquid, steam, light and a camera move, which makes differences easy to see.
Each model ran once at the default settings of the MCP generate_video tool, so length and resolution vary by model. In Studio you can pick length and resolution per model.
Results at a glance
| Model | Lab | Length | Resolution | Audio |
|---|---|---|---|---|
| Seedance 2.0 Fast | ByteDance | 8 s | 1280×720 | Yes |
| Seedance 2.5 | ByteDance | 5 s | 1280×720 | Yes |
| Kling 3.0 Pro | Kuaishou | 5 s | 1920×1080 | Yes |
| Hailuo 2.3 | MiniMax | about 6 s | 1366×768 | No |
| Wan 3.0 | Alibaba | 5 s | 1280×720 | Yes |
Each image below shows the first, middle and last frame of the clip.
Seedance 2.0 Fast

It opens wide, moves to the pour, and ends looking down on the cup from above. With eight seconds it covers the most ground of the five, and the heart pattern is crisp in the final frame. Of the five, it took "the camera pushes in" most literally.
Seedance 2.5

A bright room with greenery outside the window, then a close-up of the pour. The morning light falls naturally, and the cup and wooden counter have good texture. The last frame is still mid-pour: it spends its five seconds on the motion itself.
Kling 3.0 Pro

The only 1920×1080 clip of the five. Hard sunlight through the window and steam off the pitcher are rendered clearly. The framing is a near-static side view, with little of the requested push-in, and the pattern is only starting to form by the last frame.
Hailuo 2.3

Hailuo had the barista hold the cup while pouring, with a shallow depth of field that blurs the background. There is no audio track, and the file was the smallest of the five. The leaf pattern is simple, but the pour flows without breaks.
Wan 3.0

Wan 3.0 did the most with "a small Tokyo coffee shop": a street of wooden buildings outside the window, with people walking by. It moves from a close-up of the hands to the finished tulip pattern. It was the only one of the five with moving figures in the background.
Cost
Video is billed per clip, and the rate depends on the model, the length and the resolution. Two clips of the same length can cost differently on different models. Current rates are on the pricing page and on Studio's Generate button.
To keep the bill down, iterate on the prompt with short, low-resolution clips and save the high-resolution model for the final render.
Which model for what
- Short ads and social clips with lots of movement: Seedance 2.0 Fast.
- Product shots where light and texture matter: Seedance 2.5.
- When you need high-resolution footage: Kling 3.0 Pro, which returned 1080p.
- Lots of light, silent drafts: Hailuo 2.3.
- When the setting and atmosphere must come through: Wan 3.0.
Try it
In Studio, open Create video and switch models to send the same prompt to each one. From the API, the video generation docs cover every model. The smallest request looks like this:
curl -X POST https://api.fastmetal.ai/v1/videos/create \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "bytedance/seedance-2.0-fast", "prompt": "A barista pours latte art into a white cup, morning light, the camera slowly pushes in."}'
Other models have their own create path (Wan 3.0 is /v1/videos/wan-3-0/create); the docs list them all. Poll the returned job ID until the clip is done, then download the MP4. The full flow is in the video generation API tutorial.
FAQ
Which AI video model is best?
It depends on the job. In this one-run test, Wan 3.0 stood out for scene detail, Seedance 2.0 Fast for camera movement and Kling 3.0 Pro for resolution. Every run differs, so try two or three generations before you decide.
Can these models generate audio?
At these settings, every model except Hailuo 2.3 returned an audio track. In Studio, some models let you switch audio on or off.
Why does the same prompt give a different video every time?
Video models add randomness to every generation. To lock the composition, put an image in the start frame and animate it.
When am I charged?
When the generation request is accepted, for one clip. Checking on the job and downloading the result cost nothing extra.
Bottom line
The same prompt gave five clearly different clips. FastMetal lets you switch between all of them on one account and one prepaid balance. Create a free account and try them in Studio.