Google just rolled out Gemini Omni 1.2 Flash, and I’m genuinely impressed by how they’re positioning this. While the headline features are solid, what matters more is the direction this signals for AI video generation: faster iteration, lower costs, and better consistency are becoming table stakes.
Let me break down what actually changed, because the details reveal something important about how this technology is maturing.
The Economics Are Shifting
A 30% price drop to $0.07 per second for 720p video is meaningful, but it’s not just about raw cost. Google’s workflow design matters here. You can draft at 360p for testing, refine with multi-turn edits instead of full regenerations, then upscale winners to 1080p or 4K. This is thoughtful API design. It’s the difference between guesswork and actual production pipelines.
For developers, this changes the math on batch operations. Video generation that was economically risky at scale suddenly becomes viable. I’m already thinking about which projects in my backlog suddenly make sense with these numbers.
The 15-second cap (up from 10) might seem incremental, but it’s the difference between “proof of concept” and “actual product.” You can produce full social ads, product explainers, and branded sequences in one shot. That’s not a marginal improvement; that’s a category shift.
Consistency is Actually Hard
Here’s what I find most interesting: the improvements to reference adherence and character consistency. They’re showing examples of multi-turn edits where a model’s identity stays locked across outfit changes, camera angles, and backgrounds. That’s non-trivial.
The technical achievement here is maintaining visual coherence across extended sequences. Faces, product details, typography, and scene elements need to stay consistent across transitions and camera movements. The examples they show (the 15-second macro science explainer, the continuous Pantheon flythrough with zero jump cuts) suggest they’ve made real progress on the “coherence decay” problem that plagued earlier versions.
For developers building generative AI applications, this matters because it means less post-processing and fewer failed generations. Your error rates go down. Your time-to-production goes down.
What’s Still Missing
But let me be direct: they’re honest about the gaps. Fast sports action, rapid script changes, and complex physics are still problems. If you’re trying to generate a soccer highlight reel or a physics simulation, this isn’t your tool yet.
That honesty is actually refreshing. It tells me Google isn’t overselling, and it gives me a realistic sense of where to apply this. It’s phenomenal for static or slow-moving subjects, product-focused content, dialogue-heavy scenes, and cinematic camera work. It’s not ready for dynamic action sequences.
What Developers Should Actually Care About
The multi-turn conversational editing is the feature that changes workflows. Starting from one generated clip and swapping backgrounds, lighting, or aspect ratios without regeneration? That’s production efficiency. You’re not re-briefing the model; you’re iterating the output.
Combine that with native support in the Gemini API, Google Vids, and Gemini for Business, and you’ve got a platform where the same model works across different UX layers. That reduces friction for teams building at different scales.
The SynthID watermark they mention is also significant. It’s imperceptible to viewers but lets you prove provenance and authenticity. For enterprise applications where you need to track AI-generated content, this is important infrastructure.
The Broader Picture
I keep coming back to one thing: the economics of AI video generation are normalizing. This isn’t bleeding-edge anymore. It’s becoming another tool in the stack, and the companies shipping first at this cost level will have real competitive advantages.
Developers working on API design for multimodal applications need to think about how video generation fits into their workflows. The cost structure alone makes it worth reconsidering projects that seemed marginal six months ago.
The consistency improvements matter for quality, but the pricing matters for adoption. When you can generate 15-second video sequences for $1.05, and you can iterate without full regeneration, the barrier to experimentation drops dramatically.
This feels like the moment where video generation moves from “interesting capability” to “expected feature.”