Gemini Omni 1.1 Flash: 40s Extension, First/Last Frame, and 4K Upscaling
In this article
Google has shipped gemini-omni-1.1-flash, a production update to its native multimodal video generation and editing model. Adobe Firefly, Figma Weave, GMI Cloud, and Runway are already running it in production. The model is live in Google AI Studio, the Gemini Enterprise Agent Platform, and the Gemini app for AI Plus, Pro, and Ultra subscribers.
The release reframes video generation as a stateful editing loop. The Interactions API carries a previous_interaction_id parameter across turns; the model applies only the named change and preserves everything else, without requiring a re-upload of the prior clip. That architecture is what makes the new controls coherent rather than merely additive — see how systems-engineering thinking rivals raw scaling for the broader pattern.
Scene Extension: 10-Second Context Read, 40-Second Ceiling
The headline change is how far back the model reads when continuing a clip. Previous Omni versions referenced only the final second of footage; 1.1 reads up to 10 seconds of prior context, giving it enough signal to maintain character motion, camera trajectory, and audio coherence. Extensions run in 10-second increments to a cumulative 40 seconds, and the model generates a 3–10 second continuation per call, editing some final frames of the input to smooth the seam.
Constraints are explicit. Extension appends to the end of a clip only — no prepending, no mid-clip insertion. Uploaded input videos must be 10 seconds or shorter unless the developer is extending a model-generated video in multi-turn. New dialogue cannot be added when extending an uploaded clip where someone is already speaking; spoken dialogue is supported in multi-turn extension when passing previous_interaction_id.
First/Last Frame Interpolation and Reference Tags
Developers can pin a start frame and an end frame and have the model generate the continuous motion between them — the mechanism behind orbits, dolly-zooms, and seamless loops. The prompt syntax uses <FIRST_FRAME> and <LAST_FRAME> tags. <IMAGE_REF_N> binds a still as a style or subject reference; <VIDEO_REF_N> binds a clip for character or object consistency.
Video references accept a maximum of three clips at up to three seconds each. Audio inside a video reference is ignored. Reasoning across multiple videos simultaneously is not supported and is documented as likely to degrade output quality.
Resolution Tiers and Pricing
The response_format parameter accepts 360p, 720p (default), 1080p, and 4k. The top two are upscaled outputs rather than native renders. Google reports 360p previews generate up to 60% faster and at a third of the cost of 720p, which structures a clear draft-then-upscale production pattern.
Pricing is published. Input costs $1.50 per 1M tokens across text, image, video, and audio modalities. Output costs $9.00 per 1M text tokens and $17.50 per 1M video tokens. Video billing runs at 5,792 tokens per second of 720p, producing an effective rate of approximately $0.10 per second at standard pricing. There is no free tier and no provisioned throughput option.
| Capability | Gemini Omni 1.1 Flash | Veo 3.1 | Sora 2 / Pro | Gen-4.5 / Gen-4 Turbo |
|---|---|---|---|---|
| Max clip length | 40s cumulative via 10s extension steps | 8s base; longer via scene extension | 4–20s (Pro) | 5s or 10s |
| Extension context read | Up to 10 seconds of prior footage | Scene extension supported; window not published | Not documented | Not documented |
| First / last frame control | Yes — <FIRST_FRAME> / <LAST_FRAME> tags |
Yes — frame-specific generation | Not documented | Partial — keyframe modes on Gen-4.5 |
| Conversational / stateful editing | Yes — stateful via previous_interaction_id |
No — regenerate per call | No — regenerate per call | No — regenerate per call |
| Resolution options | 360p / 720p / 1080p / 4K (top two upscaled) | 720p / 1080p / 4K | 720p; Pro adds 1024p / 1080p | 720p |
| Cheap draft mode | Yes — 360p at up to 60% faster, ⅓ the cost of 720p | No draft tier documented | Partial — batch tier at ~50% off | Partial — Gen-4 Turbo as cost tier |
| Entry rate per second | ~$0.10 at 720p | $0.10 Fast 720p · $0.40 Standard | $0.10 · Pro from $0.30 | ~$0.05 Turbo · ~$0.12 Gen-4.5 |
| Provenance watermark | SynthID on every output | SynthID | C2PA provenance program | Not documented in API docs |
| API status | Generally available, paid tier only | Generally available, paid tier | Deprecated — Videos API shuts down Sep 24, 2026 | Generally available, self-serve |
Hard Limits
Every generated video carries SynthID watermarking — invisible to viewers but programmatically detectable. Several controls engineers expect from language-model APIs are absent: no system instructions, no temperature, no top_p, no stop sequences, and no negative prompts (Google directs negatives into the prompt text itself). Voice editing is unsupported. Audio references are unsupported. YouTube URLs cannot be used as an input source. For outputs exceeding 4 MB, the API requires delivery="uri" and polling the Files API until the file status reads ACTIVE. Editing or extending uploaded video is geofenced — unavailable in the EEA, Switzerland, and the UK. English is the only fully evaluated language; others are explicitly marked unevaluated in Google's documentation.
The Sora deprecation — the entire Videos API and all Sora snapshots retire September 24, 2026, with no named replacement — compresses the competitive landscape precisely when Omni 1.1 lands. Google's documentation had steered developers toward Veo 3.1 for scene extension and last-frame control; with Omni 1.1 now covering both, the distinction between the two Google products narrows considerably.