Gemini Omni 1.2 Flash Integration Guide: Input Modes, Parameters, and Spec-Based Billing

Model and capability boundaries

HiAPI currently offers public access to gemini-omni-1.2-flash for outputting video. It supports text-to-video, first-frame image-to-video, first/last-frame transition, and reference-image-to-video. Video editing, continuation, video or audio input are not available, and the seed parameter is not provided.

Tasks are always submitted via POST /v1/tasks; the request body uses the structure {"model":"gemini-omni-1.2-flash","input":{...},"storage":"temp"}.

Input modes and mutual exclusion rules

Choose a mode based on the material:

• Text-to-video: pass only prompt, with no image fields.
• First-frame image-to-video: pass prompt and first_frame_url, using a public image to specify the starting frame.
• First/last-frame transition: pass both first_frame_url and last_frame_url; last_frame_url depends on first_frame_url. This mode has a maximum length of 10 seconds and cannot include negative_prompt or system_prompt.
• Reference-image-to-video: pass prompt and image_urls, using 1–10 public HTTP(S) reference images to constrain style or subject. If only one image_urls image is passed, it remains a reference image, not a first frame.

image_urls is mutually exclusive with first_frame_url and last_frame_url and cannot be mixed in the same request. All image parameters accept only public HTTP(S) URLs; Base64 is not supported.

Parameters and default values

Do not assume that all modes include sound, lip sync, accurate text, or exact subject/brand consistency; none of these are guaranteed.

Two copyable POST /v1/tasks requests

The API Key in the calls is represented only by YOUR_API_KEY; no real key is written. Both requests use storage: "temp".

3-second 360p text-to-video

{
  "model": "gemini-omni-1.2-flash",
  "input": {
    "prompt": "A white paper airplane takes off from a desktop, glides past blue glass, and the camera slowly follows",
    "duration": 3,
    "resolution": "360p",
    "aspect_ratio": "16:9"
  },
  "storage": "temp"
}

10-second 720p first-frame image-to-video

The https://example.com/first-frame.png in the example is a placeholder address and must be replaced with a real, publicly accessible image before calling; it is not an actual test input.

{
  "model": "gemini-omni-1.2-flash",
  "input": {
    "prompt": "The building in the frame transitions from early morning to dusk, light moves slowly, and the camera pushes in slightly",
    "first_frame_url": "https://example.com/first-frame.png",
    "duration": 10,
    "resolution": "720p",
    "aspect_ratio": "16:9"
  },
  "storage": "temp"
}

Do not add image_urls to a first-frame image-to-video request; the two are mutually exclusive.

Submission, querying, and retry

After submission, save the returned task ID, query the task status and result via GET /v1/tasks/{id}, replacing {id} with the actual task ID. Query requests use the same API Key as submission requests.

If a network timeout occurs and the task ID has already been obtained, query the existing task first; if the task ID has not yet been obtained, retry submission with the original Idempotency-Key to recover the same task and avoid creating a duplicate task. Re-creating a task because the successful result is unsatisfactory will increase the budget.

Resolution-based billing and 10-second budget

The current default multiplier pricing is as follows:

Budget = unit price of the selected resolution × requested duration × number of items. The budget for a single 10-second item is: 360p $0.229, 720p $0.343, 1080p $0.343, 4k $0.80.

Actual billing and applicable multipliers are subject to final settlement; billing cannot be inferred from the fractional seconds in the file container. Do not assume that failed tasks are always charged or always free, and do not estimate by doubling the cost per round or assuming text request costs are constant.

In one production verification on November 11, 2026, a 3-second 360p text-to-video task took about 34 seconds and was actually charged $0.0687, consistent with that specification's unit price multiplied by the requested duration. This sample does not mean that the quality and billing of other specifications or modes have all been tested.