Google Releases Gemini Omni Flash Preview, a Multimodal Model for 720p Video Generation

Google has released Gemini Omni Flash Preview, a native multimodal model built for video generation and editing. The model accepts text, images, and video as input and outputs both text and generated video, according to Google.

What the Model Does

Gemini Omni Flash Preview generates short videos at 720p resolution with native audio, supporting text-to-video generation according to Google's model description on OpenRouter. The model's modality is listed as text+image+video → text+video, positioning it as a unified system for both understanding and generating video content rather than a separate pipeline for each task.

The model ships with a 131,072 token context window, giving it room to process substantial text, image, and video inputs in a single request.

Pricing and Availability

Gemini Omni Flash Preview is available now through OpenRouter's API as google/gemini-omni-flash-preview. Pricing is set at $1.50 per million input tokens and $9.00 per million output tokens — a notably higher output cost than Google's text-only Flash models, reflecting the compute demands of video generation.

As a "preview" release, the model carries Google's standard caveat that specifications, pricing, and behavior may change before a stable release. Google has not disclosed a training data cutoff date for this model, nor released benchmark scores against standard video-generation evaluation suites such as VBench or EvalCrafter.

What This Means

Gemini Omni Flash Preview signals Google's push to fold video generation into its mainline Gemini family rather than keeping it siloed in a separate product like Veo. Bundling text, image, and video understanding with video output in one model — accessible through the same API surface as other Gemini models — lowers the integration burden for developers who want to add video generation to existing Gemini-based applications.

The $9/M output pricing is steep relative to text generation but likely reflects the cost of producing video frames plus synchronized audio, a combination that few competitors offer in a single API call. OpenAI's Sora and other dedicated video models typically charge per-second or per-clip rather than per-token, making direct pricing comparisons difficult until more usage data becomes available.

Because this is a preview model with no independent benchmark scores yet, developers should treat output quality and reliability as unverified until third-party testing emerges. The lack of a disclosed training cutoff also means users can't yet assess how current the model's world knowledge is when generating video content tied to recent events or trends.