Gemini Omni 1.1 Flash: what shipped for video control

On August 27, 2026, Google shipped Gemini Omni 1.1 Flash — the first production-oriented update to the Omni Flash video model that launched at I/O. The official announcement and Google's X thread frame it as a control release, not a rename: extend a scene, lock a start and end frame, attach short video references, draft cheaply in 360p, then generate 1080p or 4K for the takes you keep.

That is a different job from the May 19 consumer debut we covered in Gemini Omni Flash: what shipped and what Google held back. The original Flash pitch was conversational editing and any-to-any input, with a hard 10-second clip and no scene extension in the June 30 developer API. 1.1 is the pass that tries to make those clips long enough, sharp enough, and directed enough for a production pipeline.

This article is the capability rundown: what 1.1 actually added, which limits still apply, and which of those jobs you can already run in OmniArt today.

What Gemini Omni 1.1 Flash actually shipped

Google's API id is gemini-omni-1.1-flash (preview on the Gemini Enterprise Agent Platform as gemini-omni-1.1-flash-preview). Default output is still 720p landscape; 16:9 and 9:16 remain the two aspect ratios. SynthID watermarking is still mandatory.

The withheld list from I/O did not move. Google's current docs still say voice editing is unsupported, audio-reference upload is unsupported, and avatar mode was never part of this release. 1.1 adds direction and length. It does not add the safety-held features.

Scene extension: 10 seconds of context, up to 40 seconds

Scene extension is the headline. You continue from the tail of an existing clip instead of hoping a new 10-second generation matches it.

Google's own numbers: the model now reads up to 10 seconds of prior footage — "a leap from previous models that only referenced the final second" — then appends another 3–10 seconds. Repeat that in 10-second increments until the cumulative clip hits 40 seconds. Some of the last frames of the source are rewritten so the join is not a hard cut. You cannot prepend, and you cannot extend the middle of a clip.

The prompt for an extension can be as short as "Continue the scene." or as specific as a new camera move. Google's docs are explicit that 0 seconds in a timestamped extension prompt means the start of the new segment, not the start of the original clip. Describe audio if you need the score to change; otherwise the model tries to keep it coherent.

Two constraints matter before you plan a 40-second piece around this:

• Uploaded source clips for extension must be 10 seconds or shorter, unless you are extending a clip the model already generated in the same multi-turn session.
• Extending an uploaded clip is not available in the EEA, Switzerland, or the UK. Extending a model-generated clip is available in all regions where the API itself is.

On OmniArt, Seedance 2.5 already covers a related job: extend any result of 30 seconds or less, repeatedly, up to 60 seconds, or generate 30–180 seconds in one Long Video pass. That is the workspace path for long directed video today. 1.1's 40-second Omni-native chain is the Google-surface version of the same idea.

First and last frames

Set a starting image and an ending image; Omni 1.1 generates the continuous motion between them. Google positions this for camera orbits, zoom transitions, and looping clips — the last of those by using the same image as both the first frame and the last frame.

Prompt tags bind the roles. Google documents simple FIRST_FRAME and LAST_FRAME tags inside the prompt; more complex prompts can declare sources explicitly:

[# Sources <FIRST_FRAME>@Image1 <LAST_FRAME>@Image2]

The tags stay in the prompt as written. The surrounding instruction is where you describe the move: one continuous shot, no jump cuts, what the camera does between the two stills.

On OmniArt, PixVerse V6 already exposes a transition workflow between approved start and end frames. If the brief is "get from this frame to that frame," you do not have to wait for Omni 1.1 to land as a separate model id. If the brief is "do that inside Omni Flash's conversational thread," that still lives on Google's Interactions API.

Video references, up to three seconds each

1.1 lets you drop short reference video into the same multimodal prompt as text and images. Google's public demo maps three dancer clips onto three character stills and asks for one continuous shot.

The documented ceiling is stricter than the marketing line. Video references work best for likenesses; audio inside a reference clip is ignored; you get a maximum of three clips, up to 3 seconds each. Google also says referencing or reasoning across multiple videos is not supported as a narrative compare — treat each clip as a subject or motion reference, not as a second source to edit.

That is a real expansion from the June 30 preview, which accepted one video reference and did not fully process clips longer than 3 seconds. It is still not the audio-in half of the any-to-any pitch. Sound remains something you describe in the prompt.

OmniArt's current Gemini Omni entry accepts one reference video (MP4 or MOV, 1–10 seconds) mixed with up to five reference images, and it bills from the source clip's duration. Use that for a single motion reference today. For larger reference packs, Seedance 2.5 is still the workspace model built around that job.

360p drafts and 1080p / 4K output

The resolution ladder is the other production change. Default generation is 720p. You can now request:

• 360p — Google's cheap draft tier. Claimed at up to 60% faster throughput and about one-third the cost of 720p. Flow lets you draft at 360p, then download a keeper at 720p.
• 1080p or 4K — documented as upscaled output from the 720p generation, aimed at social, digital, or broadcast finishing.

The useful workflow is the obvious one: explore camera, performance, and cut points at 360p, then spend the 1080p or 4K generation on the take you would actually ship. That is the same discipline as drafting on a fast model in OmniArt before finishing on a higher-cost one — only now it is a resolution switch on a single Omni model.

Native 4K generation is still the Veo 3.1 job in this lineup. 1.1's 4K path is a finishing step on an Omni clip, not a claim that Omni replaced Veo for large-screen delivery.

Where it is live, and what OmniArt exposes today

Google is rolling 1.1 out across its own stack on August 27:

• Google AI Studio and the Gemini API (gemini-omni-1.1-flash)
• Gemini Enterprise Agent Platform (gemini-omni-1.1-flash-preview)
• Google Flow, including start/end frames, 360p drafts, and 1080p / 4K export
• Gemini app scene extension, for Google AI Plus, Pro, and Ultra subscribers

Standard Gemini Omni generation is already on OmniArt: 720p, 3–10 seconds, native audio, up to five reference images and one reference video, Starter plan and above. OmniArt does not currently expose a separate gemini-omni-1.1-flash model id, the 40-second extension chain, the first-and-last-frame task, or the 360p / 1080p / 4K resolution ladder. Session-preserving conversational edits still require Google's Interactions API, as they did when we covered the June 30 developer API.

How to pick against Veo 3.1 and Seedance 2.5

1.1 does not retire the rest of the Google line, and it does not make Omni the only directed-video option in OmniArt. Pick per shot:

The pattern is the same one we used at I/O: new Omni releases are additions. You keep the models that already cover the brief, and you add 1.1 when the job is specifically Omni's conversational control plus a longer, sharper finish.

For prompting habits that still apply to the first 10-second beat — one clear action, sound written into the prompt, no negative-prompt field — see the Omni Flash prompt guide. For extension language on the workspace model that already does it, see the Seedance 2.5 editing and extension guide.

What to do this week on OmniArt

If you want to try Omni itself, open Gemini Omni in the video workspace and run the current 10-second, 720p surface: text, image references, or one short video reference. Compare it side by side with Veo 3.1, Seedance 2.5, PixVerse V6, Kling, and the rest of the lineup — one balance, one prompt grammar.

If the brief is specifically 1.1's new controls — 40-second Omni-native extension, first/last-frame interpolation inside Omni, or a 360p-then-4K loop — use Google AI Studio, Flow, or the Gemini API this week, and treat OmniArt as the place you already generate the rest of the project.

That split is temporary in the same way the original Flash API gap was. When 1.1's controls land as a selectable OmniArt model, they become another option in a workspace that is already producing, not a reason to wait.

FAQ

What is Gemini Omni 1.1 Flash?

Gemini Omni 1.1 Flash is Google's August 27, 2026 update to Omni Flash, a multimodal video generation and editing model. It adds scene extension, first-and-last-frame interpolation, short video references, 360p drafts, and 1080p / 4K output on top of the original conversational editing and text/image/video input.

Is Gemini Omni 1.1 Flash available on OmniArt?

Not as a separate model id. OmniArt currently serves standard Gemini Omni generation at 720p and 3–10 seconds. 1.1-specific controls — 40-second extension, first/last frames, 360p drafts, and 1080p / 4K output — are rolling out on Google AI Studio, Flow, the Gemini API, and the Gemini Enterprise Agent Platform.

How long can Omni 1.1 Flash videos be?

A single generation is still a 10-second clip. Scene extension appends 3–10 seconds at a time, using up to 10 seconds of prior footage as context, to a cumulative maximum of 40 seconds. Uploaded sources for extension must be 10 seconds or shorter.

Does Omni 1.1 Flash generate native 4K?

Google documents 1080p and 4K as upscaled output from the default 720p generation, not as a native 4K capture path. For native 4K video in OmniArt, Veo 3.1 remains the Google-lineage model built for that finish.

Did 1.1 add audio-reference upload or voice editing?

No. Google's current Omni docs still list audio-reference upload and voice editing as unsupported. Video-reference audio is ignored. Sound is still steered in the prompt, which matches what we recorded in the any-to-any input article.