What is Gemini Omni 1.2 Flash?
Google made Gemini Omni 1.2 Flash generally available on November 5, 2026. The model creates and edits video. It is also the first version of Omni that Google has opened to all Google Cloud customers.
The model's identifier is gemini-omni-1.2-flash. It accepts text, images and video, and it returns video with native audio. Google's documentation lists a 131,072-token input limit and a 260,640-token output limit.
Google introduced the Omni family in May 2026. The Next Web reports that the first Omni Flash demo at Google I/O capped clips at 10 seconds. Version 1.1 followed, and version 1.2 now raises that ceiling by half.
The headline changes are easy to list. Clips can run from three to 15 seconds. Users can edit a clip across several chat turns. Reference images hold their details more reliably, and Google cut the price by 30%.
Google positions the model for high-volume commercial work. Product videos, virtual try-ons and ads are the examples in its announcement. The company says the model is production-ready and offers Provisioned Throughput for reserved capacity.
Access spans developer tools and consumer apps. Developers reach it through the Gemini API, Google AI Studio and the Google Cloud console. Google apps reach it through Flow, Google Vids, Google Ads and, for Google AI Ultra subscribers, the Gemini app.
The Verge frames the release as an upgrade to an existing model rather than a new product line. Its report focuses on clip length, chat-based edits and smoother camera movement. That is a fair description of a release that keeps the same interface and changes its limits.
The name Omni reflects Google's plan to handle several media types in one model. The documentation shows text, image and video inputs, video output and sound generation. That is broader than a pure text-to-video tool. It helps explain why Google markets the model to ad teams, not only to filmmakers.
How long are the clips, and what does the price cut mean?
Clip length is the clearest difference between versions 1.2 and 1.1. Google says the model now generates and edits video natively for up to 15 seconds, compared with 10 seconds before. Users can also pick a duration in one-second steps between three and 15 seconds.
The price change is just as concrete. Google dropped the 720p rate by 30%, to $0.07 per second. The Next Web reports that the same rate was $0.10 per second in June. A 10-second clip that cost about $1.00 then now costs about $0.70.
Google bills video output by tokens, not by the second. At 720p, the pricing page counts 5,792 output tokens per second of video. Video output costs $12 per million tokens, which produces the advertised rate of about $0.07 per second. A 15-second 720p clip works out to about $1.04.
Other resolutions carry their own rates. Google lists about $0.023 per second for 360p and $0.104 for 1080p with upsampling. The 4K rate is about $0.209 per second with upsampling. The same 15-second shot therefore costs about $0.35 at 360p, $1.04 at 720p, $1.56 at 1080p and $3.14 at 4K.
That range rewards planning. At list rates, 1080p costs about 50% more per second than 720p. 4K costs about three times the 720p rate. A team that drafts at 720p and upscales only the final cut can avoid most of that gap.
There is no free tier for this model. The pricing page marks the free tier as unavailable and the paid tier as the only option. Input costs $1.50 per million tokens for text, images and video. Text output costs $9 per million tokens.
Google recommends a tiered workflow to control spending. Teams can draft at 360p or 720p and fix details with follow-up edits. They can then upscale only the best take to 1080p or 4K. That approach avoids paying for a full-resolution render on every attempt.
The economics matter because the clips are short. A 15-second spot is a social ad or a product loop, not a television episode. Buyers will compare the per-clip cost with the time an editor spends on revisions and approvals.
How does conversational editing change the workflow?
Conversational editing is the feature Google expects people to use most. A user can change a product, a background, the lighting, the weather or the aspect ratio without starting over. Google says these multi-turn edits work across the full 15-second clip.
The company's demo shows why that matters for advertising. One clip swaps its setting from a striped tent to a forest, then a stable, a market fair and a castle. The performances and camera framing stay in place through each turn.
The Verge describes the same capability in simpler terms. The model can make edits without re-generating the entire clip. That distinction matters when a client requests a small change late in a project.
Reference handling supports the editing workflow. Google says the model accepts character images, product photos, motion clips of up to three seconds and audio cues. It promises that faces, clothing, packaging typography and brand colours stay consistent across camera angles.
A virtual try-on example illustrates the claim. Google combined one model photo with front and back shots of three outfits. The model produced a 15-second sequence with 360-degree turns and match cuts, keeping the model's identity and each garment's cut and drape.
Audio references are not open to everyone yet. Google Cloud says audio references sit behind an allowlist. The same post does not say when that restriction will be lifted for general customers.
Editing also covers motion and speech. Google says the model handles continuous camera moves such as FPV drone flights and dramatic zoom-outs with fewer unwanted cuts. It also claims clearer lip-sync and distinct voices when several speakers share a scene.
The practical result is fewer full regenerations. Each turn costs output tokens, but a small edit should cost less than a fresh 15-second render. That makes iteration cheaper for teams that expect several rounds of client feedback.
Google's demos show the range of edits. One turns a single style frame into a four-shot science explainer with motion-tracked callouts. Another replaces a rippling octopus with a dancer whose gown copies the same fluid motion. These examples come from the vendor, so they show intent rather than measured reliability.
Why is Google cutting prices for AI video?
The 30% cut is a signal about where Google sees demand. Its launch materials emphasise volume: product listings, catalogue spins, social ads and try-on clips. Those are jobs that a brand repeats hundreds or thousands of times.
Enterprise customers want predictable capacity. Google Cloud now offers the model with Provisioned Throughput and says it is backed by the company's enterprise infrastructure and security. Photoroom's chief executive cites general availability and service-level agreements as reasons it can rely on the model in production.
Early customers describe the same bottleneck. Invideo's vice-president of engineering says the model sticks to prompts. He also says that 15 seconds gives the team a full scene in one go. Photoroom's chief executive calls 15-second videos "a top user request."
QuickFrame AI makes a similar point about television advertising. Its chief technology officer says 15-second generation at 1080p lets users create a complete connected-TV ad in a single generation. That is a customer claim, not an independent test.
Picsart sees reach as the benefit. Its head of product and partnerships says general availability lets the company bring this quality to more than 100 million Picsart creators. The figure is the company's own user count, supplied in Google's announcement.
Competition is pushing in the other direction on length. The Next Web notes that ByteDance's Seedance 2.5, unveiled in June, makes 30-second clips. Google's 15-second ceiling is generous for ads but short for narrative work.
Google also faces pressure on quality. The Next Web reports a figure from Google DeepMind's own comparisons. Testers preferred version 1.2 over 1.1 for visual consistency 62% of the time. That figure comes from the vendor's testing, so treat it as a directional signal rather than independent proof.
Lower prices and longer clips change the build-versus-buy calculation. A studio can generate more variants before committing to a shoot. A retailer can turn existing product photography into motion without booking a set or hiring a crew.
Where can you use Gemini Omni 1.2 Flash?
Developers get the model through the paid tier of the Gemini API. Google AI Studio provides a prototype environment, and the Cloud console provides the enterprise path. The documentation lists the model as gemini-omni-1.2-flash.
The API surface is narrower than Google's text models. The model card says the Interactions API is supported, while GenerateContent is not. Tool features such as function calling, grounding and code execution are also marked unsupported.
The model supports more than text-to-video. Documentation lists generation from an image, generation from references, first-and-last-frame video, video extension, video editing, upscaling and sound generation. Those features cover a short ad from first frame to final cut.
Google apps get different levels of access. Flow supports scene direction and multi-turn edits. Google Vids can generate 15-second clips at no additional cost for anyone with a Google account or Workspace subscription. Google Ads can use the model for campaign video.
The Gemini app is the slowest rollout. Google says 15-second video creation is rolling out to Google AI Ultra subscribers. Google has not announced access for the free or Pro tiers of the Gemini app.
Business users get a separate path. Google Cloud says companies can create with the model in Gemini for business and Google Vids. In Vids, teams can combine generated clips with avatars, music and voiceover for product explainers and social posts. A prompt-to-video or product-photo-to-video feature is "coming soon" to Gemini chat for business users.
Provenance tools are part of the package. Google says every output carries an imperceptible SynthID watermark. The model documentation also lists support for C2PA Content Credentials, a standard that attaches signed metadata to media files.
One limit is easy to miss. The model accepts up to 131,072 input tokens and returns up to 260,640 output tokens. Those ceilings matter for teams that send long reference sets or request several clips in one call.
What are the limitations and risks?
Google is explicit about what still needs work. The announcement names "fast sports action, rapid script changes and complex physics" as areas under development. Those are common failure points for video models, and Google does not claim they are solved.
The quality claims are mostly comparative and vendor-run. Google says reference adherence and visual consistency improved. The 62% preference figure reported by The Next Web comes from DeepMind's own comparisons. The launch materials cite no independent benchmark.
Audio references remain gated. That limits lip-sync and voice-matching workflows to approved customers for now. The Cloud post does not say when the allowlist will widen or what criteria applicants must meet.
Price is low but not trivial at scale. A 15-second 4K clip costs about $3.14 at list price. A campaign that tests 100 variants before approval would spend roughly $314 on generation alone, before editing or media spend.
Synthetic video also carries provenance risk. SynthID and C2PA help platforms identify generated media, but they depend on metadata surviving distribution. Google's documentation lists C2PA support; it does not claim that every downstream platform will check it.
The model is not a replacement for a production crew. It generates short shots with audio, and Google's own examples rely on human direction, reference photography and editing. The useful comparison is with stock footage and motion graphics, not with a full shoot.
Access is another constraint. The model has no free tier, and the Gemini app version is limited to Ultra subscribers at launch. Small teams may find the API bill manageable but the learning curve real.
The documentation also marks several familiar Gemini features as unsupported. Function calling, grounding, structured output and code execution are not available on this model. Teams that need those tools must pair Omni with a separate Gemini text model.
What should you watch next?
The first thing to watch is independent evaluation. Buyers need tests of reference adherence, lip-sync and motion across many prompts. Vendor comparisons are useful for direction, but they cannot replace a blinded test on your own material.
The second is clip length. ByteDance's Seedance 2.5 already produces 30-second clips, according to The Next Web. If rivals keep stretching duration, Google will face pressure to move beyond 15 seconds or to add stitching tools for longer sequences.
The third is price. Google cut the 720p rate from $0.10 to $0.07 per second in about five months. That comparison uses The Next Web's June figure and the current pricing page. Further cuts would change which projects become viable for small teams.
The fourth is enterprise adoption. Artlist, FLORA, invideo, Krea, Picsart, Photoroom and QuickFrame AI are named in Google's Cloud post. Their production results will show whether the model holds up outside curated demos.
The fifth is provenance policy. Watermarks and signed metadata only help if platforms, advertisers and regulators check them. Expect more scrutiny as cheap 15-second clips spread into ads and social feeds.
For now, the release is a clear step in cost and control. It gives advertisers a faster way to make short variants, and it gives Google a stronger position in a crowded video market. The missing pieces are independent proof and longer narratives.