What changed in Google’s Gemini Omni 1.2 video model? Longer clips and lower prices arrived without an independent quality test
Google made its video-generation and editing model generally available at $0.07 per second for 720p output, 30% below the previous version, and extended native clips from ten to 15 seconds. The release is broad and practical, but its quality claims rely on selected demonstrations and customer quotes rather than a published comparative benchmark.
The short version
- What happened
- Gemini Omni 1.2 Flash can generate or edit native clips from three to 15 seconds, up from ten seconds, and supports iterative conversational changes rather than requiring a complete regeneration.
- Who it affects
- Designers, marketers and educators can make longer clips and revise them in conversation, but need to count corrections and approvals rather than raw generations.
- What remains unproven
- All four linked sources are controlled by Google; there is no independent evaluation or customer-outcome study.
Keep in mind
All four linked sources are controlled by Google; there is no independent evaluation or customer-outcome study.
Selected demonstrations and endorsements do not supply a representative task count, failure denominator or confidence interval.
The version-specific model card refers important training, limitation and evaluation details to earlier family cards.
List price per generated second does not include retries, editing turns, upscaling, storage, platform fees or human review.
SynthID can support provenance but is not a universal detector or a truth guarantee.
Source analysis · 4 linked sources. Primary analysed source date: 5 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The answer: this is a wider, cheaper production release—not a verified quality breakthrough
Google has made Gemini Omni 1.2 Flash generally available for creating and editing videos with text, image, audio and video inputs. The concrete changes are easy to state. Native clips can now run from three to 15 seconds, compared with a ten-second ceiling in the previous version. Users can ask for a change to the same clip across multiple conversational turns, provide reference images or short motion and audio cues, draft at lower resolution and upscale selected outputs. The company lists a 720p price of $0.07 per generated second, 30% below Omni 1.1 Flash.
Those are material product and access changes for people who make advertisements, explainers, social clips and prototypes. They do not establish that the model is more accurate or reliable in every setting. Google's announcement shows selected examples and quotes from early customers, but it does not publish a blinded comparison, task count, failure denominator or confidence interval for its claims about stronger reference adherence, more natural speech, fewer jump cuts or better character consistency. The accurate headline is wider availability, longer clips and lower list price—not independently demonstrated superiority over rival or previous systems.[1][2][3][4]
Where the model is available—and what the price does not include
Developers can use the model through the Gemini API and Google AI Studio, while organizations can access it through Google Cloud and Gemini for business. The Cloud release describes it as generally available with Provisioned Throughput and production controls. Google Vids is rolling the model out to consumer accounts and a broad set of Workspace editions, including business, enterprise, education and nonprofit plans. The Workspace notice says the rollout began on 5 November and can take one to three days, so availability to a named user may lag the announcement.
At the listed $0.07 per second, a single 15-second 720p output has a nominal generation charge of $1.05 before any retries, extensions, editing turns, storage, upscaling, platform fees or staff time. A usable clip may require several attempts, making cost per finished asset higher than cost per generated second. Google also says people using Vids can make up to three 15-second clips a month at no additional charge, but that allowance is a product entitlement rather than evidence that professional production is free. Teams should measure completed, approved assets and correction time instead of counting raw generations.[1][3][4]
The distinction that matters
What changed—and what still needs testing
The release changes access, duration and list price; quality and safety claims need version-specific evidence.
Duration
Generate 3–15 second clips
Five seconds longer than Omni 1.1
Price
Pay $0.07 per 720p second
30% lower list price
Editing
Revise the same clip conversationally
Fewer full restarts in principle
Evidence
Inspect demos and model card
No independent comparative benchmark
Longer clips and conversational edits change the workflow
Fifteen seconds is still short, but it is long enough for a compact advertisement, product turntable, teaching vignette or social post without stitching multiple generations together. Duration can be selected in one-second increments. The model can also take a first and last frame, extend an existing clip or use visual and motion references. Google presents examples of virtual try-on, product consistency, multi-character dialogue and continuous camera moves. These illustrate intended use; they are not a representative sample of everything the system produces.
Conversational editing may have the larger workplace effect. A user can ask to replace a product, background, weather condition or camera angle while retaining the rest of a clip. That can reduce the need to restate a brief and may help teams explore variants. It also creates a version-control problem: after several edits, a reviewer needs to know which source assets, prompts, consent terms and approvals apply to the final output. A production workflow should retain the original references and every material edit rather than treating the last generated file as a self-explanatory record.[1][4]
The model card leaves important evaluation questions unanswered
Google DeepMind's model card lists text, images, audio and video as inputs and high-resolution video with audio as output. It describes broad distribution and states that the model inherits architecture and training information from the earlier Gemini Omni Flash family. For known limitations and the evaluation approach, it largely refers readers back to family-level cards rather than supplying a version-specific table of tasks, samples and results. That makes it difficult to quantify which behaviours improved and by how much.
The frontier-safety section says Google relies on evaluations of Gemini 3.1 Pro and Gemini 3.7 Flash and does not believe Omni 1.2 adds a material capability increase that would reach a tracked or critical threshold. That is a provider assessment, not an independent audit of a video model. Video-specific risks include impersonation, non-consensual editing, false documentary scenes, misleading product demonstrations and subtle changes to a person's words. A capability can matter socially at broad scale even when it does not cross a frontier-model threshold.[2]
Speech editing and virtual people require consent controls
The launch highlights clearer lip-sync, distinct speakers and the ability to condition output on voice references. The model card says video editing can change people's speech and that Google is restricting this capability while it studies safe release. Separate Cloud documentation says audio references are allowlisted. Those restrictions are important because a short, plausible clip can misrepresent a colleague, customer or public figure even when it is imperfect. A generated scene should not be described as documentary evidence or attributed to a real person without permission and verification.
Organizations need policies that travel with the tool. A real person's face, voice, clothing or performance should enter a reference workflow only with documented authority for that use and audience. Synthetic presenters and actors should be labelled where viewers could otherwise infer a real event. Advertising teams also need to verify products, prices, claims and physical behaviour in the final frame-by-frame output. Prompt compliance is not legal clearance, and a model's ability to retain packaging typography does not show that every label or disclosure is accurate.[1][2][3]
SynthID helps with provenance, but it is not a truth test
Google says Omni outputs carry an imperceptible SynthID watermark. A provider-applied signal can help platforms and investigators identify content generated by participating systems. It does not prove that a scene is true, that every edit is visible or that a file has retained its signal after conversion, cropping, screen recording or re-encoding. Nor does it identify synthetic media made with a different provider. Readers should treat detection as one piece of provenance evidence, not a universal authenticity stamp.
A responsible publishing workflow should preserve original media, generation records and explicit captions alongside any embedded signal. Platforms can test whether watermarks survive the formats they actually use and state the false-positive and false-negative behaviour of the detector. The launch page does not publish those denominators for Omni 1.2. If the model becomes common in advertising and workplace communication, practical transparency will depend as much on visible labelling, asset records and enforcement as on an invisible technical marker.[1][2][3]
What would turn a product announcement into decision-grade evidence
Creative teams should compare Omni 1.2 with their present process on a preregistered set of real briefs. Useful measures include the share of outputs accepted without correction, retries per approved clip, time to final approval, reference-identity errors, text errors, prohibited content, accessibility work and total cost. Reviewers should be blinded where feasible, and the benchmark should include difficult motion, multiple speakers, diverse skin tones, sign language, rapid action, small text and products whose shape cannot be allowed to drift.
Confidence would rise if Google or independent researchers released version-specific evaluation methods, task counts and failure categories; if outside teams reproduced the gains; and if provenance survived common transformations. It would fall if the lower per-second price were offset by more retries, if speech editing escaped consent controls or if consistency failed on ordinary briefs. For now, the release gives creators a more accessible and flexible tool. Whether it improves finished work safely and economically remains a question for measured production use, not a conclusion supplied by the launch examples.[1][2][3][4]