Gemini's Scene Extension keys each continuation to a previous interaction ID

The load-bearing line in Google's API sample is `previous_interaction_id=previous_video_interaction.id`, not the model name ^[10]. The continuation is keyed to a server-side interaction, not to a file you hand over, so in a pipeline the interaction ID becomes metadata stored beside the asset. That is a new field in somebody's schema. In the app the flow is looser: pick Create video, create or upload a clip, then describe what should happen next ^[15]. What the published material does not show is an API extension whose base is an arbitrary uploaded file, which matters the first time an editor trims the opening seconds in an NLE and hands the clip back.

Then the context arithmetic. Omni 1.1 reads up to 10 seconds of prior footage where earlier models referenced only the final second ^[4], so ten times the window ^[14]. Increments are 10 seconds to a cumulative 40 ^[1], which is four segments: an opening clip plus three continuations ^[13]. Ten seconds against a 40-second finish is a quarter of the timeline ^[16]. By segment four the model has never seen segment one. Anything that must survive the whole sequence, a logo or a product colour, has to be re-established inside each window or carried by the other controls: up to three seconds of video reference ^[7], or pinning a shot between specified first and last frames ^[11].

The sample also asks for 360p in `response_format` ^[10], which tells you the loop Google expects. Call one 720p segment one unit. Google prices 360p at a third of 720p ^[5], so four drafts come to 1.33 units against the 4 units of a full pass ^[17]. Draft and then render everything and you have spent 5.33 units for 4 units of output, a third more. Abandon the concept at draft stage and you spent 1.33 instead of 4 ^[17]. The draft resolution is an option on rejection. It only pays off if review actually rejects.

The speed figure needs the same treatment. Up to 60% faster is footnoted to system throughput of 360p against 720p ^[6], which is Google's fleet, not your wall clock. For it to land in your iteration time, generation has to be the slow step, not the queue and not the reviewer's calendar.

Google's own blog calls these updates production-ready for professional use through the Gemini API in AI Studio ^[18]. The rollout write-up on dev.to is more careful, treating "seamless" as a product claim to validate against your own footage, brand style and prompt requirements ^[8]. Both readings hold. Production-ready describes the controls, including keyframes, references and 1080p or 4K output ^[19]; it says nothing about the odds that segment three matches segment one. Google also points developers at the Gemini API, AI Studio and the Gemini Enterprise Agent Platform ^[9], while the same write-up notes no ready-made integration with any specific marketing, ecommerce or publishing system ^[12]. The continuation itself is callable. The approval step around it is still yours to build.

The honest framing for a marketing team is modular: four beats, each 10 seconds, each needing its identity cues restated, with the 4K pass reserved for the version that clears review.