Google just shipped Gemini Omni 1.1 Flash, and the headline feature is a fix for something that’s been quietly driving video creators insane: continuity. Previous models referenced only the final second of a clip when extending it, which is why AI-generated videos so often feel like jump cuts stitched together by someone who wasn’t paying attention. Omni 1.1 now reads up to 10 seconds of prior context, extends in 10-second increments up to a cumulative 40 seconds, and even edits some final frames of the input to smooth the seam. It’s the difference between a model that generates clips and one that can actually tell a continuous story.
Camera control gets a serious upgrade too. You can now pin first and last frames and let the model generate everything in between — the mechanism behind orbits, dolly-zooms, and seamless loops. Prompts bind media to roles with tags like <FIRST_FRAME>, <LAST_FRAME>, and <VIDEO_REF_N>, with video references accepting up to three clips of three seconds each for character consistency. Audio inside those references gets ignored, and reasoning across multiple videos isn’t supported yet, but for likeness work it’s a genuine step forward.
Then there’s the cost play. Google is pushing a draft-then-upscale workflow: generate 360p previews at a third of 720p cost and up to 60% faster, iterate until it works, then render once at 1080p or 4K. Actual pricing lands at $1.50 per 1M input tokens and $17.50 per 1M video output tokens, which works out to roughly $0.10 per second of 720p video. Every output carries SynthID watermarking — invisible but programmatically detectable. Adobe, Figma Weave, GMI Cloud, and Runway are already named as production users, which suggests this isn’t a beta toy.
What’s missing matters too. No system instructions, no temperature or top_p controls, no negative prompts outside the prompt text itself, no voice editing, and YouTube URLs can’t be used as sources. English is fully supported; other languages are unevaluated. That’s a meaningful constraint list for anyone expecting parity with text models. But the stateful editing via previous_interaction_id — where you pass an ID and the model applies changes while preserving everything you didn’t mention, without re-uploading video — points to where this is heading: video models you converse with, not just prompt. Runway and Pika have been circling the same idea, but Google’s distribution through the Gemini API and enterprise platform gives it reach those startups don’t have yet.
