Creating a short internal training clip or product explainer has long been a costly, multi-step process for many companies: a script, an internal or external crew, a shoot, an edit and rounds of revisions. Small changes — for example to on-screen text after legal review — often forced the whole chain to run again. Google aims to change that workflow with Gemini Omni Flash, the first model in its new "Omni" family, which is now rolling out to developers and enterprise customers via an API after debuting to consumers at I/O 2026.
What Omni Flash does
Gemini Omni Flash’s selling point is conversational editing: rather than a one-shot text-to-video prompt, the model accepts text, images and short video references and returns a finished clip with synced audio. The API brings this capability to teams that produce the most internal video, such as marketing and learning-and-development.
Replacing a five-tool pipeline with one conversation
Until now, many organizations assembled AI video workflows from multiple point tools—an LLM for script, a text-to-image model, an image-to-video model, a lip-sync tool and a voice generator—each with separate contracts, billing and data paths. Omni’s enterprise argument is unification: a single model that carries references and context across turns, reducing vendors and centralizing output monitoring and data-handling policies.
Because each instruction in a conversational session builds on the previous context, a user can relight a shot, reframe it or change wardrobe without regenerating everything from scratch — the difference between booking a reshoot and sending a note.
Multimodal references and a physics-aware world model
Omni accepts more than text prompts: users can provide up to seven reference images and up to three short video clips (each three seconds or less), and the model will carry those specifics into the generated result. A product photo or a brand logo can be dropped in as an ingredient rather than only being described in text.
Google highlights two capabilities relevant to enterprise work: a world model that understands how physical scenes behave (for example, rendering reflections in wet pavement when light rain is added) and text/logo insertion that can rewrite signage or place a company logo into a scene. Tests showed sign tracking in complex scenes was not always perfect and some text slipped back to the original language between frames, so human review remains necessary.
The interactions API and remaining limitations
Under the hood this runs on Google’s new interactions API, a stateful interface designed for multi-turn tasks rather than open-ended chat. Each turn carries the previous video and references forward, allowing edits to accumulate coherently. Developers can chain generations and store different versions to branch from later.
There are material constraints to plan for. The model’s published model card caps clips at 10 seconds; longer outputs require generating chunks and editing them together. Uploaded footage can be edited if it is 10 seconds or shorter and the user holds the rights. Google is candid that maintaining consistency across edits and rendering accurate text remain open problems.
Guardrails, watermarking and provenance
On provenance, every Omni clip carries Google’s SynthID watermark. Google is also extending C2PA Content Credentials across its generative tools and has launched an AI Content Detection API that flags AI-generated media from Google and other vendors.
Google has drawn deliberate limits: the model will not accept a still photo of a person plus an audio clip to produce lip-synced speech, a measure to limit deepfakes. It will, however, take a recording of someone talking and translate it into another language, which can be useful for localizing global training content. For regulated enterprises, these constraints and the built-in provenance features are often assets.
Pricing, technical specs and quality signals
Pricing released with the API is aggressive: Omni Flash costs $0.10 per second for generated 720p video, making a 10-second clip roughly $1. That matches Veo 3.1 Fast at 720p, is double Veo 3.1 Lite and undercuts standard Veo 3.1.
The key catch is that Omni Flash currently only generates 720p; there is no 1080p or 4K option. Veo 3.1 tiers scale up to 4K, so for premium brand work destined for large screens Veo remains relevant. Omni Flash creates clips of 3 to 10 seconds at native 720p, in landscape (16:9) or portrait (9:16). It accepts up to seven images and up to three video clips (each three seconds or less) as references. It does not yet accept audio as an input, though it generates audio alongside the produced video. Output is MP4 and every clip ships with SynthID watermarking and C2PA credentials.
Early quality signals are positive: on LMArena’s Text-to-Video Arena leaderboard, Omni Flash ranked number one with a score of 1527.
Budget implications and what’s still missing
With real pricing, iteration costs are concrete: every conversational edit is a fresh generation you pay for (roughly $1 per 10-second pass at 720p). The stateful model reduces wasted attempts because context carries across turns, so generations refine an existing take rather than starting from a blank prompt.
Omni is not alone—Veo 3.1 remains the production option for higher resolution work, and competitors from Bytedance, Alibaba and OpenAI are also targeting enterprise video budgets. Omni’s distinguishing feature is the editing capability itself: treating video as a living document rather than a one-shot render.
Author: AI & Enterprise Technology desk Date: API rollout follows the model’s consumer debut in May at I/O 2026
