By ByteBot
Google’s video generation API just changed shape. Gemini Omni Flash (model ID: gemini-omni-flash-preview) landed in the Gemini API on June 30, and it is not Veo 3.1 with a new name. It runs on the Interactions API — Google’s new primary interface for agents and models — which means you generate a video clip, then edit it in natural language turns without re-uploading anything. That is a meaningfully different workflow from every other video API on the market right now.
How the Interactions API Changes the Workflow
Every other video generation API on the market today is one-shot: you prompt, you get a clip, you tweak the prompt and start over. Runway, Veo 3.1, Sora — all of them require you to re-submit the full generation if you want to change something. Gemini Omni Flash is built differently.
The Interactions API stores conversation state server-side. Each call to client.interactions.create() returns an interaction id. Pass that ID as previous_interaction_id in your next call, and the model retrieves the video context and applies your new instruction without a full re-generation. Character consistency, camera angles, and scene continuity carry over automatically.
This is not an incremental improvement. Iterative video editing has been the core pain point for anyone building production video pipelines — you prototype a clip, the client wants a change, and you are regenerating from scratch. The Interactions API makes that loop fast and cheap.
Getting Started in 10 Lines of Python
Install the SDK:
pip install google-genaiGenerate a clip:
import base64
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-omni-flash-preview",
input="A marble rolling fast on a chain reaction track, smooth continuous shot."
)
with open("marble.mp4", "wb") as f:
f.write(base64.b64decode(interaction.output_video.data))
print(interaction.id) # save this for the next turnNow edit it conversationally — no re-upload needed:
edit = client.interactions.create(
model="gemini-omni-flash-preview",
previous_interaction_id=interaction.id,
input="Add a wooden ramp before the final loop. Keep everything else the same."
)
with open("marble_v2.mp4", "wb") as f:
f.write(base64.b64decode(edit.output_video.data))Two API calls to generate and refine. The model handles state; you pass the interaction ID and your instruction.
What It Costs and What You Can Build
Pricing is straightforward: $0.10 per second of generated 720p video. A 10-second clip costs $1.00. A 5-second animated product image costs $0.50. That is the same rate as Veo 3.1 Fast at the same resolution — you are not paying a premium for the conversational editing capability.
- Product demo pipelines: Animate a product image, iterate with the client in a feedback loop, export final clip — all in one session
- Ad creative iteration: Generate a base ad, run A/B variations through conversational edits without full re-generations
- E-learning content: Convert slide images into short explainer clips, adjust narration style through follow-up instructions
- Social content pipelines: Build Shorts/Reels pipelines where the editing loop lives in code, not in a timeline editor
There is also a clean pairing available: use Nano Banana 2 Lite to generate reference images, then animate and edit them with Omni Flash. One SDK, one billing account, one coherent pipeline from static image to polished video clip.
The Limits — Read These Before You Ship
Gemini Omni Flash is in public preview. The capability is real, but the constraints are significant enough to plan around:
- 720p only. No 1080p or 4K in preview. If your use case requires high resolution, Veo 3.1 still owns that space (up to 4K).
- 10-second cap per clip. Chain clips for longer video; the Interactions API makes sequencing straightforward but it adds latency.
- No audio editing. You can generate video with synchronized audio, but you cannot edit the audio in subsequent turns. This was deliberately held back.
- SynthID watermark — permanent and invisible. Every clip gets a pixel-level SynthID watermark at generation time. It survives re-encoding and resizing. There is no enterprise tier that removes it.
- Preview means breaking changes. The API surface will change before GA. Do not build hard dependencies on response shapes today.
The SynthID line deserves emphasis. Other providers either skip watermarking or make it optional. Google is treating it as non-negotiable at the infrastructure level. That is a reasonable policy for a technology this new, but it is a material constraint for consumer-facing applications. Your users’ output will be watermarked — plan your product accordingly.
If You’re Currently on Veo 3.1
Do not migrate your production pipeline yet. Veo 3.1 is GA, supports 4K, and has a stable API surface. Gemini Omni Flash is preview with a 720p ceiling and potential breaking changes ahead.
| Gemini Omni Flash | Veo 3.1 | |
|---|---|---|
| Editing | Multi-turn conversational | Re-prompt per iteration |
| Resolution | 720p (preview) | Up to 4K (GA) |
| Price | $0.10/sec | $0.10/sec Fast |
| Status | Public Preview | Generally Available |
The move to make right now: architect your video generation calls behind an abstraction layer. A thin wrapper that accepts a prompt and returns a video URL gives you a clean swap-in point when Omni Flash reaches GA or when the resolution ceiling lifts. These are different paradigms — Veo 3.1 for high-fidelity single generations, Omni Flash for iterative creative workflows — and your code should route between them.
Bottom Line
Gemini Omni Flash is the most architecturally interesting video API to launch in 2026. The conversational editing loop is a genuine paradigm shift, not a marketing slide. The limits are real — 720p, 10-second clips, permanent SynthID watermarks — but the core capability is production-ready for iterative workflows at $0.10/sec.
Start with the official Gemini Omni docs and the VentureBeat enterprise analysis for broader context. Build the abstraction layer now. The resolution ceiling will rise.
