Gemini Omni: video generation API

Gemini Omni generates a short video in one synchronous request: the finished clip comes back as base64 MP4 in the response body, with no task ID and no polling.

Available models

gemini-omni-flash-preview is the only Gemini model that outputs video, and unlike every other video model on HopBase it is synchronous. It returns a 3-10 second 720p MP4 and is billed by token rather than by second. See the model card for full specs.

Omni video is served by the Gemini Official Direct group — the same group and the same key as Gemini chat and Gemini image generation. There is no separate video package to buy. Confirm availability with GET /v1/models and see your effective rates in the signed-in Model Plaza.

Request parameters

There is no duration, resolution, or fps parameter. State the length you want in the prompt.

Response

The request blocks until the clip is finished, so there is no queued or processing state. On success you get HTTP 200, and the clip sits in steps[].content[] on the step whose type is model_output.

Success

{
  "status": "completed",
  "model": "gemini-omni-flash-preview",
  "steps": [{
    "type": "model_output",
    "content": [{ "type": "video", "mime_type": "video/mp4", "data": "<base64-mp4>" }]
  }],
  "usage": {
    "total_input_tokens": 20,
    "total_output_tokens": 17819,
    "output_tokens_by_modality": [{ "modality": "video", "tokens": 17376 }]
  }
}

The numbers in usage.output_tokens_by_modality are already included in total_output_tokens, so do not add them again.

The clip arrives inside the JSON body itself: there is no download link, so nothing expires. Responses are large as a result, so size your client buffers accordingly.

Failure

A failure takes one of three forms:

• A gateway validation failure returns 400 in the {"error": {"message": …}} shape; see "Common errors" below.
• Errors raised by the model itself come back in its native error body.
• A 200 without a type: "video" block (the model answered with text only, or declined). This is still billed by tokens, so check for that block before treating the call as a success.

Minimal example

curl https://api.hop-base.com/v1beta/interactions \
  -H "Authorization: Bearer sk-your-key" \
  -H "Content-Type: application/json" \
  --max-time 600 \
  -d '{
    "model": "gemini-omni-flash-preview",
    "input": "3-second video: a red ball slowly rolling on a white background"
  }'

Notes

Video generation is slow and this endpoint has no asynchronous mode — Prefer: respond-async works for Gemini image generation but not here.

• The native Interactions endpoint has no gateway-side watchdog, so set a generous client timeout — five minutes is a reasonable floor.
• Calling gemini-omni-flash-preview through /v1/chat/completions instead is capped at roughly 100 seconds per call, regardless of your client timeout; use the Interactions endpoint directly for anything that might run longer.
• HopBase deliberately does not retry a 504 on this path. Replaying a request whose outcome is unknown would risk generating and paying for the clip twice.
• The model ID carries preview. Capabilities and availability can change without notice.

The gateway currently has no uniform count or per-file byte validation for Omni image/video inputs. An unverified limit does not mean unlimited support. Do not copy GPT Image's 16-image limit or Seedance's media rules.

uri is a remote media address and data contains base64; do not use asset://, Seedance asset IDs, or local file paths. Structural acceptance does not guarantee media validation or content moderation success.

Common errors and fixes

The gateway rejects these with 400:

Interactions API only supports model gemini-omni-flash-preview
request body is missing input
missing model
request body must be JSON

Billing

Billed by token, not by second. A 200 without a video block is still billed by tokens; a request that errors or times out is not billed.

No balance is reserved up front; a 402 comes back only when your balance is zero or below. Official list prices are on the pricing page, your own rate is the one shown in the signed-in Model Plaza, and each request's actual charge is on the Usage page in the console.

Next steps

• Video generation: submit and poll video tasks for every video model
• Gemini chat: Gemini text and multimodal chat
• Concurrency, timeouts, and billing: concurrency caps, timeouts, and billing rules