Image-to-video splits on the last frame

On September 29, 2026, OpenRouter published a comparison of image-to-video models. It covers four lines: Veo 3.1, Seedance, Kling, and Grok Imagine Video, and says that is not the whole catalog. All of them use POST /api/v1/videos. What differs is the duration, resolution, frame role, audio, and price each line accepts.

Resolutions, durations, frame roles, and audio come from GET /api/v1/videos/models. Prices come from the model pages. OpenRouter checked both on September 11, 2026. The post says video prices vary with resolution, audio, and input type, so the figures are for comparison and the model page should be checked before estimating a production cost. This piece did not query the API again on September 29.

An image can be an exact frame or only a reference. frame_images places it as the first frame, the last frame, or both. input_references gives the model image, video, or audio to follow without fixing a frame. If one request sends both fields, frame_images takes precedence and the job is treated as image-to-video.

Not listed means the model-page description does not mention reference images, video, or audio. The post says that is not a test showing that references are rejected.

Veo 3.1, Fast, and Lite generate 4, 6, or 8 second clips with first-frame and last-frame control and optional audio. Veo 3.1 and Fast go up to 4K. Lite stops at 1080p. Without audio, Lite starts at $0.03 per second for 720p and Veo 3.1 starts at $0.20 per second. The posted ranges are $0.20 to $0.60 per second for Veo 3.1, $0.08 to $0.30 for Fast, and $0.03 to $0.08 for Lite. The Fast and Lite descriptions list SynthID watermarking. The post does not say that marker is on every Veo 3.1 variant.

Seedance 2.5 generates 4 to 30 second clips at 480p or 720p, accepts first and last frames, and takes up to 50 image, video, and audio reference assets. It is the only model in this comparison that produces a 20 or 30 second shot in one generation. Seedance is priced per video token, not per second. Seedance 2.5 lists $10.70 per million video tokens. The model pages state that the token count is height times width times duration times 24, divided by 1,024. The per-second figures are the starting rates shown for 480p. At $0.1028 per second, a 4 second clip starts near $0.41 and a 30 second clip near $3.08. The charge is higher at 720p or with a different input. Every Seedance model runs through hosted APIs only. OpenRouter found no published weights or model license for any of them.

Kling v3.0 Standard and Pro generate 3 to 15 second clips at 720p with first-frame and last-frame control. Audio is optional and changes the price. Standard costs $0.084 per second without audio and $0.126 with audio. Pro costs $0.112 and $0.168. Kuaishou describes Pro as the higher visual-quality tier. Kling Video O1 generates 5 or 10 second clips at $0.112 per second.

Grok Imagine Video 1.5 generates 1 to 15 second clips at 480p, 720p, or 1080p, but its supported_frame_images array lists first_frame only. The opening image can be fixed. The closing image cannot. The description says it generates sound effects, ambience, and dialogue, so the table marks audio as yes. These Grok models do not expose a generate_audio switch. The price is $0.08 to $0.25 per second plus $0.01 per input image. If the clip must end on a second exact image, the post says to use Veo, Seedance, or Kling.

Submission retries 5xx errors by default for up to about an hour. A video submission starts billable work once the server accepts it. If the job is accepted and a transient error arrives before the job ID, an automatic retry submits a second paid job and the client only sees the second one. The post therefore turns retries off for the submission. Polling is read-only and safe to repeat. The docs recommend 30 seconds between polls, and stopping on failed, cancelled, or expired.

This piece did not recheck the prices, and it did not open each model page for changes after September 11.