Gemini Omni Flash for Video Generation

On June 30, 2026, Google pushed gemini-omni-flash-preview into public preview through the Gemini API. The model is built for two workflows: generating short video from a text prompt or a still image, then refining that video through conversational back-and-forth. Per the release note, output lands in the 3 to 10 second range at 720p, and the model is wired to the Interactions API rather than a bespoke video endpoint.

The same June 30 changelog entry also promotes gemini-3.1-flash-lite-image, marketed as Nano Banana Lite, to general availability. That is Google’s low-latency lane for image generation and editing, sitting alongside the previously GA’d Nano Banana 2 (gemini-3.1-flash-image) and Nano Banana Pro (gemini-3-pro-image). Read together, Omni Flash and Nano Banana Lite fill out the fast tier of Google’s native visual model stack: one for images at ultra-low latency, one for short video with a conversational edit loop.

3-10svideo length

720poutput resolution

1 APIInteractions API surface

The conversational editing framing is the load-bearing claim. Rather than treating video generation as a one-shot render, Google routes Omni Flash through the same Interactions API surface that already handles multi-turn text, tool use, and streaming. On paper, an operator can prompt a clip, then ask for changes in follow-up turns inside the same session. The changelog does not describe how state or frames are preserved across turns, only that the workflow exists and is documented in the Omni Flash guide and model card.

Key details

Model ID: gemini-omni-flash-preview. Stage: public preview. Release date: June 30, 2026. Access surface: Interactions API.

• Supported inputs: text descriptions, still images.
• Supported outputs: 3 to 10 second videos at 720p.
• Post-generation behavior: conversational editing and refinement inside the same session.
• Companion release: gemini-3.1-flash-lite-image (Nano Banana Lite) promoted to GA, optimized for ultra-low latency image generation and editing.

Related stack context from the surrounding changelog window:

• Interactions API: generally available, recommended for all latest features and models. A breaking schema change (outputs to steps, plus response_format output config) became default May 26, 2026; legacy schema removed June 8, 2026.
• Nano Banana family GA: gemini-3.1-flash-image and gemini-3-pro-image, both GA May 28, 2026. The -preview IDs are deprecated and shut down June 25, 2026.
• Video-to-image: supported exclusively on gemini-3.1-flash-image, accepting a direct video upload or public YouTube URL alongside a text prompt.
• Veo deprecations: veo-2.0-generate-001, veo-3.0-generate-001, and veo-3.0-fast-generate-001 shut down June 30, 2026. Migration path: Veo 3.1 preview IDs or Veo 3.1 GA on the Gemini Enterprise Agent Platform.

Why it matters

Google already ships Veo, and Veo is the serious video model in the family. Omni Flash is not trying to replace it. It is a Flash-class release, which in Google’s taxonomy means fast, relatively cheap, and tuned for high call volume rather than fidelity ceilings.

What is new is packaging the fast lane as an editable conversation instead of a render queue. Most video generation today still behaves like a batch job: submit a prompt, wait, receive an artifact, throw it away if it is wrong, prompt again. The unit of work is the render. Omni Flash reframes the unit of work as a session.

Omni Flash changes the unit of work from a one-shot render to a session you can talk to and iterate on inside the same API call.

Second, Google keeps consolidating everything under the Interactions API. Speech generation, agents, tool use, and now short-form video all route through the same surface that replaced the older generateContent pattern for most flagship features. For operators, that is the real architectural signal: one client, one schema, one streaming pattern to cover text, audio, image, agent, and video. Teams still hedging on migration should read the Omni Flash launch as another reason the answer is now yes.

Third, this arrives inside a fast-tier arms race. Nano Banana Lite going GA the same day means Google’s image-editing, video-generation, and edit-loop surfaces are all running on Flash-family silicon. The bet is that most production media work does not need frontier quality, it needs iteration speed at reasonable fidelity.

How it compares

Inside Google’s own lineup, the clearest comparison is Omni Flash versus Veo 3.1. Veo is the fidelity-and-duration model, with video extension, up-to-three reference images, and first-and-last-frame conditioning. Omni Flash caps at 10 seconds, sits at 720p, and its selling point is that you can talk to it about what you got back.

Veo 3.1

• Fidelity and duration
• Video extension, reference images
• Plan the shot up front

Omni Flash

• Fast, 720p, up to 10s
• Conversational edit loop
• Iterate on what you got back

Operators picking between them are picking a workflow shape, not a quality tier. Veo asks you to plan the shot. Omni Flash asks you to iterate on it.

Against the broader field, most mainstream text-to-video systems still operate on a render-then-review loop rather than conversational refinement. What is less common is stitching the video generation call into the same multimodal session API that handles text, tools, and audio. If a competing platform ships a video model but requires a different endpoint, SDK, and streaming schema than its chat model, an operator building a product that mixes text and video ends up carrying two integrations. Omni Flash under the Interactions API is a single integration.

Who should care

• Product teams building creator tools: this lets you offer “generate a 6-second clip, then tell the model what to change” as a single feature instead of a two-stage pipeline. Rebuild the interaction pattern before you rebuild the model wiring.
• Ad-tech and social-content platforms: 720p at 3 to 10 seconds is a near-perfect fit for social ad formats and short-form feed content. The conversational edit loop shortens the human-in-the-loop cycle for creative variants.
• Platform teams already on Gemini in production: the migration cost is small if you are already on the Interactions API. If you are still on generateContent for anything shipping to users, this launch is one more reason the migration is overdue.
• Teams evaluating Veo: the June 30 Veo 3.0 shutdowns force a decision this week regardless. Add Omni Flash to the shortlist for the storyboarding or draft-variant step of the same pipeline.
• Solo builders and prototypers: a short-clip model with a conversational edit loop and a documented model card is a fast way to ship a demo. The 10-second ceiling is real, but for landing-page hero videos and tutorial teasers it is enough.
• Agent-platform builders: pair Omni Flash with the Managed Agents preview and you can put video generation inside a sandboxed autonomous workflow without leaving the Gemini API.

What changed practically

An operator now has a new model ID, gemini-omni-flash-preview, callable through the Interactions API. Inputs are text prompts and still images. Outputs are 3 to 10 second videos at 720p. After generation, the same session accepts conversational instructions to edit or refine that output. Google published a dedicated Omni Flash guide and model card.

Because Omni Flash rides on the Interactions API, the operational primitives that already work there apply: streaming, event-driven webhooks for long-running operations, and the schema that recently moved from outputs to steps. Anyone integrating Omni Flash needs to be on the current Interactions API schema, not the legacy one.

Hold the related Veo deprecation in the same mental model. Three Veo model IDs shut down June 30, 2026, the same day Omni Flash entered preview. Omni Flash is not called out as a replacement for Veo, but the timing means teams re-evaluating their video stack this week are looking at both in the same window.

Open questions

Several concrete things the release note does not answer:

• Frame rate and aspect ratio: 720p is stated, but frame rate, aspect ratio, and portrait support are not addressed.
• Audio: the note says nothing about audio in the generated video.
• Conversational edit scope: what does the model preserve across turns? This is the single most important thing for anyone building on Omni Flash, and the changelog does not answer it.
• Rate limits and quotas: not addressed, and for a preview-tier video model, caps determine whether a real product can be built on it.
• Watermarking and provenance: no mention of SynthID or provenance tagging. Compliance teams need to confirm before shipping.
• Enterprise availability: the note does not say whether Omni Flash is or will be available on the Gemini Enterprise Agent Platform.
• Preview-to-GA timeline: -preview means the usual caveats. Assume a rename and a schema check are coming.

Source: ai.google.dev