Google opens Gemini Omni Flash to developers through Gemini API and Vertex AI

• Google opened Gemini Omni Flash across creator and developer surfaces, with Gemini app, Flow, AI Studio, the Gemini API, and Gemini Enterprise Agent Platform named in Google's access note.
• Omni Flash is a conversational video model for generation and editing, and Google's main Omni thread shows the core pitch: camera changes, environment edits, sketch animation, style transfer, and concept visualization.
• The API shape is the bookmarkable part: Google's docs use the gemini-omni-flash-preview model slug through the Interactions API, while Google's access note confirms the API is part of the rollout.
• The preview caveats are real: Google's main Omni thread says audio references are part of the product story, while Google's own model cards split on whether audio input is supported.
• The Aug. 7 builder post is demo-led, and fofrAI's Processing experiment is the most community-native example in the evidence pool: generative Processing outputs turned into underwater creatures.

Google's builder post includes five concrete workflows, from voice-driven weather changes to doodle-guided motion. The Gemini API guide exposes the stateful editing primitive, previous_interaction_id, plus a task enum for text_to_video, image_to_video, reference_to_video, and edit. The awkward doc split: Google's Gemini API model card lists text, image, and video input, while the Enterprise model card says audio input is not supported.

Google's public post names five places where Omni is available:

• Gemini app.
• Google Flow.
• Google AI Studio.
• Gemini API.
• Gemini Enterprise Agent Platform.

The Google Cloud path is an API product, not just a UI launch. The Cloud text-to-video docs route synchronous and asynchronous requests through https://aiplatform.googleapis.com/v1beta1/projects/PROJECT_ID/locations/global/interactions, with global as the location and gemini-omni-flash-preview as the accepted model ID.

The API uses an interaction object rather than the older Veo-style generation call. In the Gemini Omni API guide, each turn creates an interaction, and follow-up edits can pass previous_interaction_id so the model keeps video state across turns.

Core mechanics from the docs:

• Model slug: gemini-omni-flash-preview.
• REST endpoint: POST /v1beta/interactions for the Gemini API.
• SDK convenience field: interaction.output_video.
• Raw REST output: video appears inside the steps array.
• Aspect ratios: 16:9 by default, with 9:16 supported.
• Task hints: text_to_video, image_to_video, reference_to_video, and edit.
• Uploaded video editing: the docs use the Files API, then pass the uploaded URI into an interaction.

The Gemini API model card labels Omni Flash as preview, lists a 1,048,576-token context window, and describes 3-to-10-second 720p output at 24 FPS.

Google's demos are best read as a capability map:

1. Perspective switching: the city-scene demo that Google's main Omni thread led with moves around one subject across about 20 viewpoints.
2. Weather and season edits: Carlos Santana changed lighting, clouds, rain sounds, leaves, and snow with voice input, according to Google's weather edit.
3. Doodle-guided motion: Google's sketch demo says Pan turned everyday-object sketches, including an espresso cup, lemon, and match, into animated clips.
4. Style transfer: Jerrod Lew's clip moved from live action to anime and other styles without breaking the walking motion, according to Google's style demo.
5. Concept visualization: Google's Hyperagent demo lists a landscape proposal, dashboard explainer, and gamified to-do list.

fofrAI's clip is the least enterprise-coded demo here: fofrAI's Processing experiment used Omni to turn Processing outputs into underwater creatures.

The Gemini API pricing page puts Omni Flash Preview on the paid tier only: no free-tier access, $1.50 per 1M input tokens, $9 per 1M text output tokens, and $17.50 per 1M video output tokens. The same page says video billing uses 5,792 output tokens per second of 720p video, equivalent to $0.10 per second under standard pricing.

A Google AI Developers Forum quota thread shows the preview mechanics developers were still hitting in July: one user reported limit: 0 for video generation, and a Google forum account pointed them to the rate-limit increase form.

Google's story and Google's docs do not line up cleanly on audio. The builder post says Omni can generate from text, image, video, or audio references, and Google's main Omni thread repeats that framing.

The implementation docs narrow that claim. The Gemini API model card lists supported input data as text, image, and video, while the Enterprise model card marks audio input as not supported. Google's June Cloud launch post adds the missing roadmap note: audio references, video references, last frame, scene extension, and higher resolutions for the Enterprise Agent Platform API were "available soon" in the June 30 Cloud announcement.

A Google AI Developers Forum complaint described a different preview failure mode: a harmless sci-fi prompt allegedly triggered a safety block, consumed credits, and did not give a real initial-frame workflow.

Google's Aug. 7 post is a showcase, not a benchmark. Independent tests give a sharper split between general video quality and controlled correctness.

• Design Arena's July evaluation ranked Gemini Omni Flash first on its Video Arena, 102 Elo points ahead, and seven places above Veo 3 Fast.
• K-Dense's 35-case science-video benchmark found the model visually strong but unreliable on text, physics, and factual correctness for unattended scientific or educational videos.
• Gemini Lab's edit-boundary test found simple lighting changes landed, while "pan the camera just 30 degrees to the right" often turned into a push-in or a replaced shot.