Google released Gemini Omni 1.2 Flash on November 5, 2026. The stable Gemini API model is gemini-omni-1.2-flash, with native audio, generation and editing up to 15 seconds, and lower video-output pricing. It is a video model; descriptions of ordinary chatbot response speed do not explain this release.
This guide checks the announcement, model specification and developer guide on November 6. AniKuku has not run paid generations or independently benchmarked the quality improvements Google demonstrates.
What is available, and where?
The Gemini Developer API offers 1.2 on its paid tier, with no API free tier. Google AI Studio is a prototyping entry point; API billing and account eligibility still apply. Flow, Vids, the Gemini app and Google Cloud have their own rollout, subscription and regional conditions. A feature shown in one product is not automatically exposed by another API.
Check the supported-region list and your project's active quotas. The list includes Japan, South Korea and Taiwan; mainland China and Hong Kong are not listed. Availability is not a promise that every account can immediately run every task.
Generation, interpolation, editing and extension have different limits
| Task in the Gemini API | Current documented limit |
|---|---|
| New video generation | 3–15 seconds, in whole-second increments; default 10 seconds |
| First and last frame interpolation | 3–10 seconds; two frames with a longer duration return an error |
| Editing an existing video | Input video at most 15 seconds; conversational editing through the Interactions API |
| Extending a video | Add 3–10 seconds per turn; final sequence at most 30 seconds |
| Extension context | Uses the final 10 seconds of the input; appends at the end only |
Output is 24 fps. The default is 720p, with 360p also available; 1080p and 4K are upscaled output, not a claim of native 4K generation. Landscape 16:9 and portrait 9:16 are supported.
For references, all durations accept up to six images. A new 11–15-second generation accepts one reference video of up to three seconds; a 3–10-second generation accepts up to three such videos. Reference-video audio is ignored, and video references are not supported for editing or extension. Reference images can be used for those tasks.
The current API does not accept uploaded audio or audio references, and does not support voice editing. Google's product announcement shows broader audio-conditioned examples; do not turn those demonstrations into an API input promise. The guide also excludes editing or extending uploaded videos in the EEA, Switzerland and UK, while allowing those operations on model-generated videos.
API pricing: what does “about $0.07 per second” include?
The Standard Gemini API price is $12 per million video-output tokens, plus $1.50 per million text/image/video input tokens and $9 per million text-output tokens, including thinking. These are USD API rates, not Gemini subscription credits.
At 720p, Google counts 5,792 video tokens per second:
5,792 × $12 / 1,000,000 = $0.069504 per output second
| Output duration at 720p | Video-output charge only |
|---|---|
| 5 seconds | $0.34752 |
| 10 seconds | $0.69504 |
| 15 seconds | $1.04256 |
These are calculated examples, not measured invoices or an all-inclusive quote. Input and other output tokens, edits and additional attempts need their own budget. At the published token rates, 360p output is about $0.023/s, upscaled 1080p $0.104/s and upscaled 4K $0.209/s. Check actual usage and current billing before a batch.
For a hypothetical 100 five-second 720p generations, video output alone is $34.752. If only 50 clips are usable, that component becomes $0.69504 per accepted clip. The 50% acceptance rate is a teaching assumption, not a model benchmark; editing time and other charges are excluded.
Start with one short shot, then edit one thing
The documented endpoint is POST https://generativelanguage.googleapis.com/v1beta/interactions. This request body follows its video contract; it is an untested integration example requiring your own authorized paid project and authentication:
{
"model": "gemini-omni-1.2-flash",
"input": "One continuous shot. An adult courier in a rust-red jacket holds a blue parcel at a rainy station. The courier looks toward the arriving train and stays still. Soft rain and distant station ambience. No cuts.",
"response_format": {
"type": "video",
"duration": "5s",
"resolution": "720p",
"aspect_ratio": "16:9"
}
}
Review the returned video before continuing. To test an edit, use the returned interaction ID as previous_interaction_id and request a single change, such as clear weather, while retaining the person, parcel and camera position. Do not set store=false if you need later edits through that ID. For large responses, follow the guide's URI-delivery flow rather than assuming every result is inline base64.
Define the role of each image in natural language. Deprecated <FIRST_FRAME> and similar control tags should not be pasted into prompts. Google advises trying natural-language task selection first; video_config.task is a fallback with stricter constraints. English is fully supported; the guide says other languages have not been evaluated. Test dialogue and captions in the intended language rather than assuming the five localized versions of this article prove five-language generation quality.
For animation, inspect identity, costume, prop contact, camera continuity and lip-sync at the start, middle and end. Change one variable per attempt and record request settings, usage and acceptance reasons. A longer shot is not automatically a better storyboard unit; our script-to-storyboard workflow explains the handoff. For a different hosted model's billing structure, see the Seedance API cost guide, without transferring its rates to Omni.
Migrating from Omni 1.1
The release notes give a four-week transition: from December 3, 2026, requests to gemini-omni-1.1-flash automatically route to 1.2 and use 1.2 rates. Before that date, explicitly test the new ID with saved assets and prompts. Compare task behavior, reference handling, dialogue, delivery and actual usage; automatic routing is not an output-identical guarantee. The older gemini-omni-flash-preview is shut down.
The model comparison guide provides the wider selection framework. This article documents Google's API; it does not announce an AniKuku integration. The cover is original concept art, not Gemini-generated test footage.