OpenAPI AI Video Guide

AI Video

AI Video uses ai_video. Select a model from generation_models[]; do not create model-specific endpoints or content types.

Gemini Omni Flash

{  "type": "ai_video",  "prompt": "Move slowly through a futuristic city on a rainy night, with neon reflections and natural ambient audio",  "aspect_ratio": "16:9",  "video_generation": {    "model": "gemini-omni-flash",    "task": "text_to_video",    "resolution": "720p",    "duration_seconds": 8,    "native_audio": true  }}

Gemini Omni Flash supports iterate conversational editing. The client never manages interaction IDs.

Veo 3.1 Fast

{  "type": "ai_video",  "prompt": "A paper airplane moves through a bright modern library with cinematic camera motion",  "aspect_ratio": "16:9",  "video_generation": {    "model": "veo-3.1-fast",    "task": "text_to_video",    "resolution": "720p",    "duration_seconds": 8,    "native_audio": true  }}

Veo 3.1 Standard

{  "type": "ai_video",  "prompt": "A coastal research station at sunrise, slow push-in camera, realistic ambient audio",  "aspect_ratio": "16:9",  "video_generation": {    "model": "veo-3.1-standard",    "task": "text_to_video",    "resolution": "720p",    "duration_seconds": 8,    "native_audio": true  }}

Veo does not support conversational editing. The current Public OpenAPI supports text-to-video only; first frames, last frames, and attachments remain unavailable until the catalog says otherwise.

Choose AI Video for motion-first work

AI Video is a fit for a short visual moment, product atmosphere, cinematic B-roll, or a concept where movement and sound carry the point. Write one coherent prompt that names the subject, action, setting, camera behavior, and sound intent. For a long instructional narrative with controlled explanation, use Explainer Video or Vector Animation instead of asking a short video to carry the whole lesson.

Select a model from current capability data

Read /catalog immediately before building the video_generation object. The selected model determines the supported task, ratio, resolution, duration, native audio, and whether a completed result can be iterated. Keep fields nested under video_generation in OpenAPI; the flat MCP argument shape does not apply here. Do not send first-frame, last-frame, or attachment fields unless the catalog explicitly marks them supported.

Create the job, poll its returned job_id, and only then read the completed creation detail. Gemini Omni Flash can use iterate only when the result exposes that action. Veo does not support it, so a material revision requires a new request with a changed prompt.

Catalog for model-specific capability fields
Creations & Jobs for status polling and result retrieval
Editing & Errors for the iteration boundary