Gemini Omni 1.2 Flash Released: 15-Second Long Video and 30% Price Cut

Google DeepMind product manager Anish Nangia announced the launch of Gemini Omni 1.2 Flash, featuring native 15-second video generation, stronger reference adherence, and multi-turn conversational editing, along with a 30% price reduction. The model is now generally available through the Gemini API, Google Cloud, and Gemini for business, and has been integrated into consumer products including Flow, Google Vids, and the Gemini app.

For teams producing video content, the highlight of this update isn't just "longer" — it's "more controllable." The cost of making revisions after generation has been significantly reduced.

Core Capability: From 10 Seconds to 15 Seconds, with Per-Second Duration Control

Gemini Omni 1.2 Flash extends native generation duration from 10 seconds in version 1.1 to 15 seconds, and supports specifying durations from 3 to 15 seconds in 1-second increments. Official use cases include complete storytelling, product explainer videos, and producing a full social ad spot in a single generation.

The 15-second length may look like just a numerical change, but it matters significantly for advertising and e-commerce scenarios: it's enough to accommodate a complete "opening—showcase—closing" structure, reducing the coherence loss caused by stitching multiple segments together.

Cost Structure: 30% Lower Unit Price, Workflow Savings Too

On pricing, Gemini Omni 1.2 Flash is 30% cheaper than Omni 1.1, with 720p video priced at $0.07 per second.

Beyond the unit price, the official team emphasizes that the workflow itself also reduces costs: you can quickly generate drafts at 360p or 720p, fix details with a single follow-up edit when issues arise rather than regenerating the entire segment, and finally upscale the satisfactory result to 1080p or 4K. This approach is especially friendly to teams producing ad variants at high frequency.

Reference Adherence: More Stable Characters, Products, Fonts, and Brand Colors

1.2 Flash supports combining character images, product images, style images, short motion clips up to 3 seconds, and audio prompts to guide generation, while maintaining these details throughout the entire video. According to the official description, faces, clothing, packaging fonts, and brand colors remain consistent across different camera angles.

One official example: using a model photo paired with front and back product images of three outfits to generate a 15-second virtual try-on sequence, covering a 360-degree turn and match cuts, while keeping the model's identity and each garment's cut, drape, and back details consistent.

In a comparison demo, the team tested 1.1 and 1.2 with the same static studio-lit poster image and a five-layer animation prompt: 1.1 Flash's fidelity to the reference image fell short of 1.2, which built up the poster layer by layer on a stage with a dusty pink background, preserving material, shadow, and typography details — ribbed glass, warm filaments, faceted crystals, and spiraling cable lettering — through to the final frame.

Reference adherence refers to how faithfully a model preserves input reference materials during generation, and is one of the key metrics for measuring the controllability of video generation models. Unlike text prompts that merely describe "a model in a red dress," reference images give the model identity, material, and typography information directly in visual form. The challenge is that the model must continuously lock onto these features across different camera angles, lighting, and occlusion changes without identity drift or texture repainting. Common failure modes include faces gradually "switching people," packaging fonts being respelled, and brand colors shifting. Evaluation typically runs multiple generations with the same set of reference images, comparing identity similarity and fine detail consistency between first and last frames.

Multi-Turn Conversational Editing: Change the Background Without Starting Over

1.2 Flash supports multi-turn conversational editing across the entire 15-second video, allowing you to replace products or backgrounds, adjust lighting and weather, and switch aspect ratios turn by turn without regenerating from scratch.

An official demo starts from an already-generated clip and, through multiple rounds of prompts, sequentially replaces the background with a striped tent, an oak forest, a stable with a white horse, a marketplace, and a stone castle, while keeping the two characters' performances, armor reflections, and camera composition unchanged. For teams needing to produce ad variants for different audiences and placements, this eliminates the repeated cycle of "re-briefing and regenerating."

Speech, Camera Work, and Consistency

On speech, dialogue generated by 1.2 Flash is smoother and more natural, with clearer lip sync, and supports multi-speaker conditioning. An official example uses a greenhouse scene image paired with four characters, mouth and 1-second voice references; the model routes each sampled voice to the correct character and maintains expressive lip sync and natural comedic timing throughout the 15-second animated scene.

On camera control, the model can execute complex continuous camera movements, such as large-scale FPV fly-throughs, dramatic pull-backs, and orbits. The team says unwanted jump cuts are significantly reduced, scale is more stable, and the overall feel is closer to cinema. One example, guided by two architectural reference images, completes a continuous FPV drone flight from a plaza through a bronze gate and up along the Pantheon dome to the oculus, with zero jump cuts throughout.

On consistency, the team emphasizes reduced drift and deformation of characters, objects, and scenes from the first frame to the last, whether interpolating between first and last keyframes or extending clips to build longer sequences — products, characters, props, and environments all remain stable.

First-and-last keyframe interpolation and clip extension are two approaches to extending the temporal dimension in video generation. The former takes a start frame and an end frame and has the model fill in the intermediate motion, ideal for precisely controlling the starting and ending composition; the latter conditions on an existing clip and continues writing forward, maintaining camera and subject continuity. The shared challenge is consistency: interpolation can produce unnatural mid-segment motion or abrupt rhythm changes, while extension may cause gradual drift in character appearance and scene lighting due to error accumulation. The team's emphasis that 1.2 Flash reduces drift in first/last-frame and extension scenarios essentially means improved error control and spatiotemporal modeling in long-sequence generation.

Availability and Enterprise-Grade Assurance

Gemini Omni is now generally available, backed by Google Cloud's enterprise-grade infrastructure and security capabilities, with support for Provisioned Throughput and outputs carrying an imperceptible SynthID watermark.

Here's how to use it through each entry point:

• Developers: Prototype in Google AI Studio; refer to the developer documentation and pricing page.
• Enterprises: Try it in the Google Cloud console and build enterprise solutions through the API.
• General users: Direct scenes and multi-turn edits in Flow; generate 15-second clips for free in Google Vids with a Google account or Google Workspace subscription; generate 15-second videos in the Gemini app (rolling out to Google AI Ultra subscribers); produce video ads at scale in Google Ads.

SynthID watermarking is an AI-generated content provenance technology proposed by Google DeepMind. Rather than overlaying a visible mark in the corner of the frame, it embeds statistical patterns directly into pixel or audio signals — imperceptible to the human eye, yet detectable with high probability by a dedicated detector. For advertising and brand content, the significance of SynthID is that once a generated video spreads across social platforms or ad networks, platforms and reviewers can determine whether it was produced by a Google model, providing verifiable clues in deepfake, copyright dispute, and ad compliance scenarios. This also means that when teams perform downstream re-editing, compression, or transcoding, the watermark's survival rate becomes a variable that content governance needs to monitor.

Provisioned Throughput is a capacity assurance mechanism Google Cloud offers for generative models: customers reserve quota at an agreed throughput rather than sharing an on-demand pool with other tenants. For video generation — where individual requests take a long time and concurrent draft and upscale tasks interleave — on-demand calls can easily queue during peak hours, directly affecting ad scheduling. Provisioned Throughput turns latency and availability from "best effort" into a predictable service level, typically used alongside enterprise-grade quota management, auditing, and data residency requirements, and is a prerequisite for moving a model from experimentation into production pipelines.

Current Limitations

The team also acknowledges that it is still improving scenarios such as fast sports action, rapid script changes, and complex physics. In other words, 1.2 Flash is positioned as a large-scale video production tool for enterprises, developers, and creators — not a universal solution for all dynamic scenes. For teams that need high-frequency output and prioritize brand consistency and editing efficiency, this combination of capabilities is quite targeted.