Gemini Omni 1.2 Flash

google/gemini-omni-1.2-flash generates 3 to 15 second clips with audio. Later requests can edit the same clip in natural language, extend it, or upscale it to a higher resolution. Requests go to /v1/video/generations.

For a cheap workflow, draft at 360p or 720p, edit the clip you like and upscale last. Generating at a high resolution from the start costs more and takes longer.

Google fully supports English prompts. Prompts in other languages can work, but results may vary.

First request

curl "$LLMTR_BASE_URL/v1/video/generations" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  --max-time 620 \
  -d '{
    "model": "google/gemini-omni-1.2-flash",
    "prompt": "A paper airplane glides across a quiet library, soft daylight, single continuous shot.",
    "duration_seconds": 5,
    "resolution": "360p",
    "aspect_ratio": "16:9"
  }'

The video comes back as a base64 MP4 in data[0].b64_json; url is null on this model.

{
  "id": "ChAxMjM0NTY3ODkw",
  "object": "video.generation",
  "model": "google/gemini-omni-1.2-flash",
  "status": "completed",
  "store": false,
  "data": [
    { "url": null, "b64_json": "AAAAIGZ0eXBpc29t...", "mime_type": "video/mp4", "duration": 5 }
  ],
  "usage": {
    "input_tokens": 100,
    "output_tokens": 10000,
    "reasoning_tokens": 500,
    "total_tokens": 10600,
    "duration": 5,
    "resolution": "360p"
  }
}

The numbers in this example only show the format.

Parameters

Field	Description
model	google/gemini-omni-1.2-flash.
prompt	Text description of the video. Required except for upscaling.
duration_seconds	Whole number from 3 to 15. Required. At most 10 with both a first and a last frame.
resolution	360p, 720p, 1080p or 4k. Without it, 720p; when continuing an earlier video, that video's resolution.
aspect_ratio	16:9 or 9:16. Without it, the ratio is inferred from the prompt. When you send your own video the ratio comes from the video and this field is ignored.
media	A list of images or a video. Each item has type and url (see below).
image_url	A single first frame, the same as first_frame in media.
store	When true, the request and the video are stored by Google for 7 days and the video can be edited later. Default false.
previous_id	The id of the video to continue. Send it with store: true.
task	upscale only.
max_wait_seconds	The longest to wait for the video. On this model both the default and the ceiling are 600 seconds.
max_budget_usd	The most you allow for this request. If the estimate is higher, the request answers 400 budget_too_low without starting.

negative_prompt is ignored and named in the answer's llmtr_dropped_parameters field. Say what to avoid in the prompt, for example "show no text".

Video generation can take minutes. If a request passes 15 seconds, the 200 status is sent right away and blank lines flow until the answer arrives; JSON parsers skip them, so there is nothing to configure. An error after that point comes in the body as an error field, so check for error in the answer as well.

Video from images

media[].url takes base64 data URLs only. Images must be PNG, JPEG or WebP; a remote address answers 400 unsupported_input.

type	What it does	Limit
first_frame	Fixes the first frame	1
last_frame	Fixes the last frame, needs a first_frame	1
reference_image	Carries a person, product or style without fixing a frame	6
video	Edits or extends your own video, MP4	1

Frames and reference images cannot share a request. A video cannot be combined with other media.

{
  "model": "google/gemini-omni-1.2-flash",
  "prompt": "A smooth transition from the first image to the second image.",
  "duration_seconds": 5,
  "resolution": "720p",
  "media": [
    { "type": "first_frame", "url": "data:image/png;base64,iVBORw0KGgo..." },
    { "type": "last_frame", "url": "data:image/png;base64,iVBORw0KGgo..." }
  ]
}

Google may refuse images of recognizable people.

Editing and extending

To edit a video later, send store: true when you generate it. Use the id from that answer on the next request, which needs store: true as well.

{
  "model": "google/gemini-omni-1.2-flash",
  "previous_id": "ChAxMjM0NTY3ODkw",
  "store": true,
  "prompt": "Make the paper airplane bright red. Keep everything else the same.",
  "duration_seconds": 5
}

The prompt decides whether this is an edit or an extension. To extend, say so: Extend this video: the airplane lands on an open book.

• Editing: the output is as long as the source. Send the source's length as duration_seconds.
• Extending: adds 3 to 10 seconds per turn. duration_seconds is the time to add. The answer returns the joined video, data[0].duration gives the total length, and only the added seconds are billed.
• media cannot be sent with previous_id.
• previous_id works only on the account that generated the video. Another account's id, or an id older than 7 days, answers 404 not_found.

When you send your own video in media as video, the prompt decides the same way and store is not needed.

Upscaling

Upscale an earlier video with task: "upscale". This request takes no media, and prompt and aspect_ratio are ignored; resolution is required and duration_seconds must be the source video's length.

{
  "model": "google/gemini-omni-1.2-flash",
  "previous_id": "ChAxMjM0NTY3ODkw",
  "store": true,
  "task": "upscale",
  "resolution": "1080p",
  "duration_seconds": 5
}

Source	Target	Tokens/second	$/second
360p	720p	2,896	0.034752
360p	1080p	5,792	0.069504
720p	1080p	2,896	0.034752
720p	4k	11,584	0.139008

There is no direct upscale from 360p to 4K or from 1080p to 4K. Videos longer than 15 seconds cannot be upscaled. Upscaling bills no input tokens and no thinking.

Pricing

No platform margin is added to the model price. The bill comes from the token counts in the answer.

Item	Price ($/1M tokens)
Input (text, image, video)	1.50
Text output and thinking	9.00
Video output	12.00

Video output tokens follow from resolution and length:

Resolution	Tokens/second	$/second	10 seconds	15 seconds
360p	1,931	0.023172	0.23172	0.34758
720p	5,792	0.069504	0.69504	1.04256
1080p	8,688	0.104256	1.04256	1.56384
4k	17,376	0.208512	2.08512	3.12768

Input and thinking tokens are added to these. The model thinks before it renders and no parameter turns that off; the real figure is in usage.reasoning_tokens.

• Each image counts for at most 1,120 input tokens.
• On edits and extensions the earlier video is input too: at most 5,792 tokens per second.

There is no cache discount or batch price.

Before the request starts, the most it can cost is set aside from your balance; once the video is delivered the actual tokens are billed and the rest is released. If generation fails or the wait runs out, the whole amount is released.

Errors

Status	Type	When
400	unsupported_input	A duration, resolution, ratio, media type or combination of fields is not valid on this model. The message says which.
400	content_filter	Google's content policy blocked the request. Change the prompt or the media.
400	invalid_request	The model could not process the request. Check the duration, the resolution and the media.
400	budget_too_low	The estimate is above max_budget_usd.
404	not_found	The previous_id is not on this account or has expired.
503	model_unavailable	The model is unavailable right now. Try again later.
504	provider_timeout	The video did not finish within max_wait_seconds. You are not charged.

Failed requests are not charged to your account.

Limits

• Tool calling, structured output, system instructions and temperature are not supported. The model cannot be called on /v1/chat/completions.
• Audio is output only; you cannot send audio input or an audio reference.
• Generated videos carry an invisible SynthID watermark.

Data

Requests are sent to Google. With store: true, the request and the video are stored by Google for 7 days; without it, storage is turned off. LLMTR does not keep the prompt or the video.

For the general video guide, see Video Generation.