Gemini API Video Generation: Complete Developer Guide (2026)
The Gemini API now supports video generation and conversational video editing through the Interactions API — the endpoint that powers gemini omni, Google DeepMind’s any-input-to-video model. Since June 30, 2026, the model gemini-omni-flash-preview lets developers generate video from text, compose it from up to 7 reference images, edit existing clips, and run multi-turn conversational edits — the only major developer API with native multi-turn video editing at launch.
Authentication and Endpoint
Get an API key from Google AI Studio and install the SDK (Python: pip install google-generativeai; Node: npm install @google/generative-ai). Video generation does not use generateContent — it routes through a dedicated Interactions endpoint:
POST https://generativelanguage.googleapis.com/v1beta/interactionsEach successful response returns a session ID; pass it back in later calls to continue conversational editing within the same video context.
Four Task Parameters
| Task | Input | Use case |
|---|---|---|
text_to_video | Text prompt | Generate video from a description |
image_to_video | Up to 7 images + prompt | Compose video from reference images |
reference_to_video | Existing video clip + instruction | Edit an existing clip |
edit | Session ID + new instruction | Continue conversational editing |
The edit task retains full context across turns — a change in turn 3 does not undo context from turn 1 — which is what makes the API uniquely suited to iterative workflows. Note that editing or replacing speech/audio in existing videos is deliberately withheld pending safety testing (voice-cloning risk).
Output and Pricing
All responses return a base64-encoded 720p MP4, up to 10 seconds, in 9:16 or 16:9, with an automatic SynthID watermark (survives compression and re-encoding) plus C2PA Content Credentials. Pricing is $0.10 per second of output ($0.50 for 5s, $1.00 for the 10s max), billed per generated second, so errored generations aren’t charged. YouTube Shorts users reach the same model free; the API tier is for developers building applications.
How the Model Works
gemini-omni-flash-preview fuses four Google AI systems: the Gemini reasoning engine (decides what to change vs. preserve), Veo rendering (720p frames), Genie world simulation (physics-aware motion), and Nano Banana frame-level editing. The reasoning layer is what enables conversational editing — it reads “make the sky more dramatic” as a targeted background change, not a full regeneration, unlike one-shot generators such as Sora or Seedance 2.0. Full details on the Google DeepMind Gemini page.
For longer narrative productions, Google Flow wraps the same API in a filmmaker UI. See the related Conversational Editing and Google Flow guides in this cluster. Permitted uses are covered in Google’s Generative AI Use Policy.
FAQ
What model ID do I use?
gemini-omni-flash-preview — the first public model in the Gemini Omni family. A higher-tier Omni Pro model is planned with no confirmed date.
Is this the same endpoint as generateContent?
No. Video generation uses the Interactions API (/v1beta/interactions); the request and response shapes differ.
Can I generate videos longer than 10 seconds?
Not currently. 10 seconds is the max for gemini-omni-flash-preview — a deployment-tier choice; longer clips are expected in future versions.
Where are the official docs?
ai.google.dev/gemini-api/docs/omni for video specifics, and ai.google.dev/gemini-api/docs for the full API reference.