Google Releases Gemini Omni 1.2 Flash: 15-Second AI Video at $0.07 per Second

Google launches another new video generation model, breaking through on both price and duration

According to MarkTechPost, Google has officially released Gemini Omni 1.2 Flash, a multimodal model for video generation and editing. Compared with previous versions, the new model advances on three dimensions simultaneously: native support for 15-second video clips, the introduction of multi-turn conversational editing, and a 30% price cut for 720p video generation to $0.07 per second.

These specifications carry notable significance in the current AI video generation landscape. A native duration of 15 seconds means the model can cover a complete narrative segment in a single generation, rather than stitching together multiple short clips to fill the runtime; multi-turn conversational editing advances the "generate—revise" loop from repeatedly retrying prompts to an interaction closer to human collaboration; and the lower price directly affects whether developers can embed video generation capabilities into high-frequency product workflows.

15-second native video: from "clip" to "paragraph"

Video generation models have long been constrained by duration. Most models can only reliably output a few seconds of footage per call, and beyond that length they tend to suffer from subject drift, scene jumps, or discontinuous motion. Gemini Omni 1.2 Flash raising the native duration to 15 seconds essentially represents a stronger answer on temporal consistency.

Another value of 15 seconds lies in the shift of application scenarios. Clips of a few seconds suit dynamic posters or looping backgrounds, while 15 seconds is already close to the typical starting length of a single piece of content on short-video platforms—enough to carry a complete product showcase, a plot twist, or a set of instructional demonstrations. For content creators, this means the model's output upgrades from "raw material" to "a usable first draft."

Temporal consistency is a core challenge in video generation: the model must not only ensure the quality of individual frames but also keep the appearance, position, and motion of the same subject coherent across consecutive frames. Common failure modes include subject drift (a person's face or clothing changing between frames), scene jumps (the background suddenly switching), and motion breaks (unnatural movement trajectories). When diffusion models generate frame by frame or clip by clip, errors accumulate over time, and the longer the clip, the harder it is to maintain stability. Therefore, every few seconds added to native duration often requires simultaneous improvements in temporal attention mechanisms, long-context modeling, and training data organization—which is why 15-second native generation is technically more demanding than stitching together multiple short clips.

Multi-turn conversational editing: treating video as a conversational object

The new model supports multi-turn conversational editing, a capability that is easy to underestimate in this update. The traditional video generation workflow is "write a prompt—generate—dissatisfied—revise the prompt—generate again," where each revision is almost a complete restart, making it hard for creators to precisely control a specific local area.

Conversational editing changes this loop. Users can continuously make revision requests on the same video, and the model makes incremental adjustments while preserving existing content. This interaction pattern aligns with a direction already validated in the image editing field, but it is more difficult to implement on temporal data like video, because any local change may affect the consistency of preceding and following frames. Google bringing this capability down to the Flash series indicates that its engineering maturity in temporal modeling has made notable progress.

Conversational editing is fundamentally different from traditional prompt-based regeneration. The latter is an independent sampling on each call; the model does not retain the previous result, and users can only indirectly influence the output by adjusting the text description, making precise local modification nearly impossible. Conversational editing, by contrast, requires the model to take the previous video as conditional input, understand the user's incremental instructions (such as "change the background to night" or "make the person turn around"), and modify only the relevant region while keeping the rest unchanged. This involves three capabilities: instruction understanding, region localization, and temporal consistency preservation—harder on video than on images, because local changes must blend naturally with adjacent frames. The image editing field already has validated paradigms such as InstructPix2Pix, while the engineering difficulty on the video side is an order of magnitude higher.

$0.07 per second: the industry implications of a 30% price cut

The most direct figure in this release is the price. 720p video generation drops to $0.07 per second, a 30% reduction from before. Calculated at the full 15-second length, the generation cost of a single video is just over one dollar.

The impact of this price point on the developer ecosystem may be more far-reaching than the specification improvements. Once the unit cost of video generation enters the "batch trial-and-error" range, product design logic changes: applications can generate multiple candidate versions first and then select, rather than carefully making just one call; marketing tools can automatically generate short videos for each product without having to make trade-offs over cost. The Flash series has always been positioned for low latency and low cost, and Google clearly hopes to use price leverage to push video generation into higher-frequency usage scenarios.

As a side note, $0.07 corresponds to the 720p specification. The pricing strategy for higher resolutions was not disclosed in the original report, and actual deployment still requires checking the official billing documentation.

Per-second billing is a common pricing method in the commercialization of video generation. Compared with per-call or per-token billing, it is closer to video duration, the core cost variable. Per-second pricing makes it easy for developers to estimate the cost of a single piece of content: 15 seconds × $0.07 is about $1.05, and batch-generating 100 candidate videos costs about one hundred dollars. This predictability is important for product design—teams can use it to judge whether a feature is worth launching, and how to price or set limits on the user side. Resolution is usually a key tiering dimension for pricing, and the unit price difference between 720p and 1080p or 4K is often significant, so comparing prices across vendors at different resolutions requires aligning specifications first.

Competitive landscape: video generation enters the cost-effectiveness competition stage

The release cadence of Gemini Omni 1.2 Flash reflects a shift in the competitive focus of the AI video generation landscape. Early competition centered on visual quality and the wow factor of creativity; today, more and more updates revolve around duration, controllability, and unit cost.

These three directions correspond precisely to the three thresholds of video generation commercialization: duration determines whether it can stand alone as a finished piece, controllability determines whether it can integrate into professional workflows, and cost determines whether it can be invoked at scale. Google's choice to advance all three on the Flash product line shows a clear strategic intent—not to win on demo videos, but to compete for developers' default invocation entry point.

For developers and content teams in China, the iteration of such models means technology choices need to be re-evaluated more frequently. The cost curve of video generation is still falling rapidly; usage that requires careful calculation today may become routine calls that can be opened up after the next version is released.

Conclusion

The keywords for Gemini Omni 1.2 Flash can be summarized as: 15-second native duration, multi-turn conversational editing, and 720p pricing at $0.07 per second. Together, these three capabilities point in the same direction—turning video generation from an occasionally invoked demo tool into infrastructure that can be embedded in daily production workflows. For teams focused on landing AI applications, it is worth rethinking both the cost model and the product interaction.