Google launches Gemini Omni 1.1 Flash, turning its video model from a generator into a steerable tool

The new model (gemini-omni-1.1-flash) is available now through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Pricing is $1.50 per million input tokens of text, image, video or audio, $9 per million output text tokens, or $17.50 per million video tokens. In practice, one second of 720p video consumes 5,792 tokens, bringing the cost to about $0.10 (approximately 37 agorot) per second at standard pricing. Every generated video carries an invisible SynthID watermark that can be detected programmatically.

The primary change in version 1.1 is Scene Extension, which allows the model to read up to 10 seconds of prior context instead of a single frame in earlier versions. Execution proceeds in 10-second segments up to a total of 40 seconds, with each call producing a continuation of 3 to 10 seconds. The limitations are explicit: additional content can be appended only at the end of a clip, not at the beginning or middle; video input is limited to 10 seconds unless it is a continuation of a multi-step generated video; and new dialogue cannot be added when extending an uploaded video that already contains speech—dialogue is supported only in a multi-step continuation via previous_interaction_id.

Now users can supply a first frame and a last frame and have the model generate the continuous video between them, enabling orbits, dolly-zoom and smooth loops. Prompts link media to roles using the tags FIRST_FRAME, LAST_FRAME, IMAGE_REF_N and VIDEO_REF_N. Video reference accepts up to three clips of three seconds each, works best for character imagination, and largely ignores any audio they contain; thinking across multiple videos is not supported and may degrade the result.

Cost control is achieved through a draft-then-upscale workflow. The resolution parameter in response_format accepts 360p, 720p (default), 1080p and 4k, with the latter two undergoing upscaling. Google states that 360p views run up to 60 % faster and cost one-third of the 720p price, based on system throughput. This makes cheap iterative drafts at 360p followed by a single final render at higher resolution the intended production pattern.

Editing is performed via the Interactions API in a stateful manner: previous_interaction_id is passed, and the model applies the change while preserving unspecified parts, without re-uploading the previous video. Missing capabilities include system instructions, temperature, top_p, stop sequences or negative prompts (negatives must be entered in the text prompt); voice editing is not supported; audio references are not supported; YouTube URLs are not accepted as input; English is fully supported while other languages have not been tested. Files larger than 4 MB require delivery="uri" and polling the Files API until status ACTIVE.

Adobe (Firefly), Figma Weave, GMI Cloud and Runway are already listed as customers running Omni Flash in production. The model is also available in Google Flow for AI Plus, Pro and Ultra subscribers, with Scene Extension in the Gemini app.