Gemini Omni 1.2 Flash Makes AI Video Worth a Production Pilot
Google put 15-second video generation and conversational editing into general availability at a lower API price, but teams should judge it by approved assets, not cheap seconds.
What shipped
Google made Gemini Omni 1.2 Flash generally available on November 5 through the Gemini API and Google Cloud. It generates and edits video with audio up to 15 seconds, accepts reference assets and supports multi-turn conversational changes.
The published 720p price is $0.07 per second, 30% below version 1.1. Teams can draft at 360p or 720p and upscale selected outputs to 1080p or 4K.
Why it matters
General availability, provisioned throughput and a published rate make this possible to budget and operate rather than merely demonstrate. Longer clips and edits to an existing result could reduce complete regenerations.
The catch is that Google published no named external benchmark for the consistency claims. Quality still has to be proven on a team’s own products, characters, languages and brand constraints.
What we would do
Virtual Arc would first run a shadow evaluation on a fixed set of real jobs. We would measure retries, rejected requests, human review time, latency, reference fidelity and cost per approved asset.
If it beats the existing workflow, it earns repeatable short-form traffic behind a provider-neutral layer. Open-ended user generation and brand-critical output stay behind tighter controls.
Google’s 30% price cut is not the reason to move; general availability plus 15-second generation and full-clip conversational editing are. Together they turn Gemini Omni 1.2 Flash from a demo endpoint into a credible production primitive for catalog video, ad variants and product explainers. But we would not replace an existing media stack during launch week. Google published no named external benchmark for its quality claims, and cheap seconds can still produce expensive assets if references drift, moderation rejects valid jobs or each revision forces another full render. We would run it beside the current workflow on a fixed set of real assets, measuring retries, review time, latency, identity and product fidelity, and cost per approved clip rather than cost per generated second. If it wins there, we would route repeatable short-form jobs to it behind a provider-neutral layer, preserve prompts and provenance, and require human approval before publication. For open-ended user generation, multilingual dialogue or brand-critical work, we would wait for our own failure data before widening access.
Turn insight into a working system.
We build a provider-neutral video generation pipeline that versions prompts and reference assets, shadow-tests models, records retries and approvals, and preserves provenance for every output. We add review gates and route each job by quality, latency and cost per approved asset. Talk to us if short-form video is already part of your product or operations.