On August 27, Google shipped its video model Gemini Omni 1.1 Flash to developers. Google AI Studio opened access the same day, the API for the Gemini Enterprise agent platform was wired in simultaneously, and Google Flow and the Gemini app went live for Plus, Pro and Ultra subscribers. Google added a line of positioning to its blog post for this release:
"Omni 1.1 is production-ready for professional use via the Gemini API."
The continuation window grows from one second to ten
The most substantive change is in scene continuity. The version from late June only read the last second of a clip when continuing it, which handed the model information roughly equivalent to a single still frame: it had no way to tell which direction the camera was pushing toward, or how far along a character's hand motion had gotten. The resulting second segment often looked like it had been shot on a different set entirely — cross-scene character consistency couldn't be held together, something Google itself admitted in the documentation for the previous version.
1.1 extends that window to the first ten seconds. Continuation still proceeds in ten-second segments, but the cumulative length can now stack up to 40 seconds. A paired new feature lets creators specify both the first and last frame, turning transitions from "generate first, then pick the best one" into "pin down both ends and let the model fill in the middle." Anyone who works in storyboards will recognize this logic: once the keyframes are locked, the in-between motion actually has something to extrapolate from.
Draft mode, 4K upscaling, and 3-second video references
The remaining changes are aimed squarely at workflow: a new 360p draft mode renders up to 60% faster than 720p, useful for quickly screening composition options; once a cut is locked in, it can be upscaled to 4K; and on the multimodal input side, video references have been added — a single reference clip can run up to three seconds, where previously only text and images could be fed in. Final output comes in 1080p and 4K tiers.
The video reference feature changes how input gets expressed. Previously, if you wanted the model to learn a particular camera move, the only option was a text description like "slow leftward pan with slight handheld shake," and whether the model actually picked up on it was mostly a matter of luck. Now you can simply drop in a three-second sample clip. Three seconds isn't enough to hold an entire scene, but it's enough to hold one complete camera move or one action loop — for series work that needs a consistent visual language, that's considerably less effort than writing ten lines of prompt text. The 360p draft tier is designed to pair with it: use low resolution first to lock down the structural elements — pacing, composition, camera movement — then switch to high resolution to fill in detail, without having to pay full price on every single round.
What does a 40-second clip actually cost
Google didn't publish a separate per-second price for each resolution this time, but it did leave a conversion basis in the pricing documentation: video output is billed by token, 720p works out to 5,792 tokens per second, and standard-tier video output costs $17.50 per million tokens — which backs out to roughly $0.10 per second. That figure is unchanged from the version at the end of June; it hasn't moved in two months.
Working from the 720p rate, a coherent 40-second clip comes out to roughly $4. The math for draft mode runs separately: if you first use 360p to narrow ten composition options down to three, then take those three through to 720p or 1080p for the final cut, the same budget buys you two extra rounds of testing camera language. Google hasn't published the actual per-unit price for 1080p or 4K; based on the token-billing structure, higher resolution should mean more tokens per unit of time and a higher price, but the exact multiplier will have to wait for the documentation to be filled in.
The duration ceiling keeps getting pushed back
Strung together, the Omni product line's cadence is fairly clear. The preview version from late June capped a single segment at 10 seconds, priced at $0.1 per second, with the limitation spelled out plainly. Two months later, the length of a single segment hasn't changed, but it can now be chained three more times, quadrupling total length, while the context window has grown tenfold. Google's approach is segment stitching: rather than piling on the length of a single generation, it leans on each segment being accurately picked up by the next to build up total duration. That approach is friendlier to VRAM and inference cost, but the trade-off is that seam stability has to be propped up by the context window — and growing that window from one second to ten is exactly what's patching the hole.
Forty seconds isn't enough to cut a full ad, but for the ten-to-twenty-second short videos that dominate social platforms, one segment is enough and two leaves room to edit. When Google wrote "production-ready for professional use" into its blog post, it was betting on exactly this kind of work — high volume, modest value per clip, but with a baseline requirement for continuity.
Sources: Google Developer Blog, Google DeepMind, CocoLoop, Gemini API pricing documentation; per-second conversion verified against the 720p rate of 5,792 tokens per second and standard-tier video output pricing of $17.50 per million tokens.