Two stills into Kling 3: 5-second reveal, silent vs audio cost

A product reveal from two ChatGPT Image 2.5 stills (square, medium, $0.0165 each) and a 5-second Kling 3 clip costs $0.73 silent or $1.08 with generated audio on Sume. Kling 3 accepts both first_frame and last_frame, so the clip is a transition between the two images.

The Kling rates are $0.14 per second silent and $0.21 per second with audio, from the Video API catalog; the stills are repository estimates.

How the two frames are sent

Each entry in frame_images carries its own frame_type. Kling 3 lists ratios 16:9, 9:16 and 1:1, resolutions 720p and 1080p, and 4 to 15 seconds, so the square stills fit the 1:1 ratio. It lists no input_references, which means style references are not an option on this model.

{
  "model": "kling-3",
  "prompt": "The box opens and the product rises into frame",
  "frame_images": [
    {"type": "image_url", "image_url": {"url": "https://media.sume.com/img/A/0.png"}, "frame_type": "first_frame"},
    {"type": "image_url", "image_url": {"url": "https://media.sume.com/img/B/0.png"}, "frame_type": "last_frame"}
  ],
  "aspect_ratio": "1:1",
  "resolution": "720p",
  "duration": 5,
  "generate_audio": false
}

Cost at 5 and 10 seconds

The stills are a fixed pair; the clip length and the audio flag set the rest.

When the audio flag is worth it

Audio costs $0.35 more over a 5-second clip, an extra $0.07 per second. For a product reveal that you will lay a music track over later, the silent version is the better value, and generate_audio can be set to false explicitly. For a clip that must carry its own ambient sound, such as a box opening or liquid pouring, the audio version saves a separate sound step.

Decide per clip, not per project. Generate the silent clip first, and add generated audio only for the shots where silence looks wrong in the edit.

• Silent clips pair well with a music bed from a separate step.
• Audio clips are better for foley-heavy moments.
• Compare a 5-second pair before committing to a 10-second one.

Making the end frame match

The last frame should be the same object in a new state, not a new scene. Generate the second still by editing the first with input_references and a prompt that changes one thing, so the identity carries over. An edit sends the first still as an input image, so the second still costs slightly more than the table assumes.

Set generate_audio on purpose. The default is the model's audio capability, and Kling 3 can produce sound, so leaving it unset can bill the higher rate.

For a batch, the stills are the cheaper part. Ten reveals cost about $0.33 in stills at medium, while the clips run $7.00 silent or $10.50 with audio at 5 seconds each, so spend your effort on the prompt and the end frame rather than the image tier.