Google made video coherence ten times cheaper in five days
Gemini Omni 1.1 Flash reads ten seconds of a scene before extending it, up from one, and the API now spends up to 88 percent fewer tokens on long video.
Until last week, a generative video model extending a shot could remember one second of what came before.
Google shipped both halves of its video stack in five days, and both moved along the same axis. On 27 August, Gemini Omni 1.1 Flash went production-ready with scene extension that reads up to ten seconds of prior footage. On 1 September, Logan Kilpatrick introduced Agentic Video in the Gemini API, which he says cuts token consumption on long videos by up to 88 percent while improving quality.
Generation got ten times more context to look back on. Understanding got roughly eight times cheaper to look at all.
Neither announcement is about fidelity. Both are about memory.
Why the look-back number beats the resolution number
A generative video model extending a shot is doing what a language model does with a context window, and until now it was doing it with a window one second wide.
One second is enough to keep the colour of a coat (and, on the evidence of the past two years, not always that). It is not enough to remember that the man in the coat was mid-sentence, or which way the camera had been drifting for the last eight seconds.
So ten seconds is a different problem. The visible result is that clips can now be extended in ten-second increments to a cumulative forty, and that the model will take up to three seconds of reference video to hold a character's look across shots.
Forty seconds is still short. But it is four calls instead of forty, and the failure mode of generative video has always been the seam.
The pricing tells the same story from the other end
Omni 1.1 offers 360p drafts that Google says run up to 60 percent faster and cost a third of the standard 720p, with upscaling to 1080p or 4K saved for a final pass.
Draft cheap, render once (the oldest idea in the business). Anyone who has waited on a render farm will recognise that workflow, and it took generative video roughly two years to arrive where post-production settled in the 1990s.
What the binding constraint actually was
Not how good a single second looks. The price of remembering the previous one, which makes both of these releases cost engineering wearing a capability costume.
We would expect the headline number in the next round of video model announcements to be seconds of maintained context rather than pixels. If a lab leads with 8K instead, we called this wrong.
So what is the longest coherent shot you actually need? For most advertising, under fifteen seconds. For a title sequence, sixty, and nobody is cutting a feature this way yet.
Google has quietly moved the ceiling to the point where the first of those is solved and the second is one increment away. The announcement that gets there will probably read as boring.
