Gemini Omni Flash 1.1: The Complete Guide — Access via Google Flow, Capabilities, Benchmarks & How to Use It

Google shipped Gemini Omni Flash 1.1 (gemini-omni-1.1-flash) on 27 August 2026 as the first generally available (GA) version of the Omni video line. The preview endpoint gemini-omni-flash-preview remains live until 30 September 2026, then it is deprecated.

Think of Omni Flash 1.1 as Nano Banana for video: you create and edit video through conversation, from any combination of text, images, video and audio, while Gemini’s world knowledge keeps physics, history and story logic coherent. Every output is watermarked with SynthID + C2PA Content Credentials.

This guide covers everything you asked for: how to access it via Google Flow, every new and existing capability, where it ranks against Seedance, Veo, Kling and Sora, pricing, and exactly how to use it in Flow and via the API.

Quick links you will reuse:

• Google Flow (creator studio): flow.google — also at labs.google/fx/tools/flow
• Gemini app (scene extension): gemini.google.com
• Google AI Studio (developer API): aistudio.google.com/prompts/new_chat?model=gemini-omni-1.1-flash
• Enterprise API: Gemini Enterprise Agent Platform → gemini-omni-1.1-flash
• Docs: ai.google.dev/gemini-api/docs/omni · Cookbook: Get_started_Omni.ipynb

At a glance — what changed in 1.1

Important fine print Google documents: The 40 seconds is stitched from four 10-second generations chained via previous_interaction_id, and 1080p/4K are upscales on top of a 720p render, not native 4K generation. Plan budgets and client promises around that.

All capabilities — the full Omni 1.1 toolset

1. Native multimodality — text + image + video + audio in one context

Omni is a transformer with native multimodal support. You can mix in one prompt:

• Text — scene brief, style, camera, audio direction
• Images — 1 to 10 images as first frame, last frame, or subject/style references (FIRST_FRAME, LAST_FRAME, IMAGE_REF_0 tags)
• Video — up to 3 seconds of reference video per generation for character/environment consistency
• Audio — generated natively (dialogue, SFX, music) with each video; input audio references are in the schema but currently not processed by the model — verify in docs before building on it

For business use, this means: drop in a product photo, a 3-second clip of a dancer, a voice style description and a text brief, and Omni fuses them into one coherent shot.

2. Conversational editing (stateful, multi-turn)

The Interactions API keeps the video live in conversation. You generate once, then say “make the violin invisible, keep everything else the same” — no re-upload, no timeline.

• Use previous_interaction_id to chain turns
• Each edit re-renders the clip — describe only the one change + append “Keep everything else the same.”
• One change per turn is roughly 2× more reliable than stacking three edits in one sentence
• After ~4 edits drift creeps in — re-anchor with “keep all character details exactly as they are, only change X”

This is Omni’s biggest differentiator. Seedance and Kling regenerate with references; Omni mutates your source clip.

3. Scene extension — grow a clip to 40 seconds

Continue an existing video without a jump cut:

• Works in 10-second increments to a cumulative 40 seconds
• Each step reads the last 10 seconds of prior footage (vs 1 second before) for character, motion and story continuity
• Append-only — cannot prepend or stretch the middle; “after 2s” in an extension means 2 seconds into the new footage
• Can attach a new reference image during extension to introduce a character

from google import genai

client = genai.Client()

# First generation
first = client.interactions.create(
    model="gemini-omni-1.1-flash",
    input="A woman playing violin on a rooftop at sunset, single continuous shot, no cuts, ambient city sound.",
    response_format={"resolution": "720p"}
)

# Extend it — reads last 10s of `first`
extended = client.interactions.create(
    model="gemini-omni-1.1-flash",
    previous_interaction_id=first.id,
    input=[{"type": "text", "text": "Continue the scene. Camera slowly pulls back as she keeps playing, warm evening light."}],
    response_format={"resolution": "720p"},
)

Flow exposes the same as Extend on any clip; the Gemini app also has scene extension for Plus/Pro/Ultra subscribers.

4. First and last frame interpolation

Give the model a starting still and an ending still — it generates continuous motion between them. Perfect for:

• Camera orbits around a product or person
• Whip-pan transitions
• Seamless loops (use same image as both first and last frame)

In the API, tag your uploads as FIRST_FRAME and LAST_FRAME. In Flow, drop both images into the Frames to Video / First & Last Frame tool.

5. Video references (up to 3 seconds)

Attach up to 3 seconds of video to preserve a character’s look, a dance move, or an environment. Example from Google:

Use the three uploaded videos of dancers and replace them with the provided characters. The dog from dog.png does the classical dance from dance3.mp4, the octopus does hip hop from dance1.mp4 — one continuous shot, no cuts.

Tag as VIDEO_REF_0, VIDEO_REF_1, etc.

6. Draft in 360p, finish in 720p / 1080p / 4K

Google’s own guidance: draft at 360p, lock the prompt and keyframes, upscale the keeper once at 1080p/4K. A 40-second scene costs roughly $1.20 as 360p draft vs $12.00 at 4K at list price.

7. World knowledge, physics & text sync

• Real-world reasoning: Gemini’s knowledge of history, science and culture → stories that follow real logic (e.g., claymation explainer of protein folding, accurate hippocampus explainer)
• Physics intuition: gravity, inertia, fluid dynamics — marbles on a chain-reaction track actually roll
• Text/action sync: on-screen text renders legibly and in time with action — lower-thirds for 26 alphabet items at 24 FPS, title cards, signage

8. Safety & provenance

• Invisible SynthID watermark + C2PA Content Credentials on every creation/edit (Flow, Gemini app, YouTube)
• Content safety filters on input and output
• Some limits to know: voice/speech editing is restricted for now, uploading images with minors is blocked in EEA/UK/CH, uploaded-video editing blocked in EEA/UK/CH, multi-video reasoning not supported, YouTube video as source not supported

Where it stands — benchmarks vs other leaders

A note on “OmniPlot”

There is no official benchmark called “OmniPlot” for video models. If you meant Artificial Analysis Video Arena, Design Arena Video Arena, VBench or MovieGenBench — Omni Flash is ranked on all of those. Below is where it stands on each.

Artificial Analysis Video Arena (human blind pairwise votes — the most cited arena)

These are rolling Elo scores from millions of blind votes. Gaps of a few points are within the ±7–14 confidence interval — treat them as ties.

Text-to-Video

Image-to-Video

Video Editing

Bottom line: Omni Flash and Seedance 2.0 / MiniMax H3 are co-leaders inside statistical error. Omni leads text-to-video narrowly; Seedance leads image-to-video and editing narrowly. The real difference is not Elo — it is features.

Different arenas publish different score scales (AA vs designarena.ai). Absolute numbers differ; relative order is consistent: Omni + Seedance on top, Veo/Kling/Sora a tier below on crowd preference.

Design Arena — Video Arena (internal human ratings)

Google cites on deepmind.google/models/gemini-omni:

• Video Editing: #1 on Overall Preference & Instruction Following (504 side-by-side comparisons)
• Text-to-Video (MovieGenBench, 1,003 prompts): #1 on Overall Preference & Instruction Following
• Reference-to-Video (468 examples): #1 on Overall Preference & Speech Adherence
• Image-to-Video (VBench I2V, 355 pairs): tied for #1 among Gemini Omni Flash, Grok Imagine and Kling

Google’s internal benchmarks (vs leading video models)

Omni Flash beats other leading models on internal blind tests for Overall Preference and Instruction Following — strongest in Complex Edits, Object Editing, Visual Effects, Physics; weakest in Sound & Speech Editing (ranked 3rd) per Artificial Analysis breakdown.

How it compares feature-for-feature

Rule of thumb from creators:

• Iterate by talking to a clip, render synced text, need world-knowledge stories → Gemini Omni 1.1 Flash
• Hard action, stunts, keep a face locked across shots → Seedance 2.0
• Need native 4K, 21:9, 15s+ continuous, or feed real audio → Kling 3.0 / Sora 2 Pro / MiniMax H3
• Best pipeline (many pros): storyboard & edit with Omni 1.1 at 360p → upscale keeper at 1080p/4K; use Seedance/Kling for hero shots that need native 4K

How to access Gemini Omni 1.1 Flash

Option 1 — Google Flow (recommended for creators)

What is Google Flow? Google’s AI creative studio at flow.google (also labs.google/fx/tools/flow). It wraps Veo 3.1, Gemini Omni Flash, Nano Banana Pro and Flow Tools (text-to-video, frames-to-video, ingredients-to-video, Scenebuilder, video extension, characters, avatars) in one workspace.

Who can use it:

*Prices as listed on one.google.com/about/google-ai-plans and support.google.com/flow/answer/16526234 at time of writing — check your account, they shift.

Credit costs in Flow (per generation):

Option 2 — Gemini app

Scene extension is also live in the Gemini app at gemini.google.com for AI Plus, Pro and Ultra subscribers globally. You can extend a clip conversationally right in chat — useful for quick drafts on mobile, then upscale in Flow.

Option 3 — YouTube Shorts & YouTube Create (free entry point)

Gemini Omni Flash is rolling out free on YouTube Shorts Remix and the YouTube Create app (18+). Scope is limited to remix/edit workflows inside YouTube — not the full Flow multimodal studio. Good for trying the editing feel without paying.

Option 4 — Google AI Studio + Gemini API (for developers)

• AI Studio: aistudio.google.com → gemini-omni-1.1-flash
• API model ID: gemini-omni-1.1-flash (paid tier only — no free tier)
• Enterprise: Gemini Enterprise Agent Platform — same model ID

API pricing (paid tier, per 1M tokens):

Token math (with audio):

• 360p: 1,931 tokens/sec → ~$0.03/sec
• 720p: 5,792 tokens/sec → ~$0.10/sec
• 1080p: 8,688 tokens/sec → ~$0.15/sec
• 4K: 17,376 tokens/sec → ~$0.30/sec

So a 40-second scene = $1.20 at 360p draft → $4.00 at 720p → $12.00 at 4K. Input images cost ~1,120 tokens each, audio seconds ~32 tokens — negligible next to video tokens.

How to use Gemini Omni 1.1 — step by step

In Google Flow (no coding)

Step 1 — Open Flow and pick the model

Go to flow.google and sign in with the Google account that has your AI Plus/Pro/Ultra subscription. Click New project (or open an existing one). In the generation panel at the bottom-right, click the model selector (defaults to Veo 3.1 — Lite) and choose Gemini Omni Flash / Gemini Omni 1.1 Flash. Confirm:

• Aspect ratio: 16:9 for web/TV, 9:16 for Reels/Shorts/Stories
• Duration: 4 / 6 / 8 / 10 seconds — 8s is the sweet spot for most work
• Resolution: start at 360p for drafts, switch to 720p when happy

If you do not see Omni immediately, the rollout may still be propagating in your region — check back in 24–48 hours. You need a personal Google AI plan; paid Workspace alone may show Upgrade.

Step 2 — Write your prompt like a shot brief

Do not write vague prose. Write a short shot brief in this order:

1. Subject + action (who, what they do — required)
2. Scene (where, time of day, weather — optional)
3. Camera + light (shot size, movement, lens, light — optional)
4. Audio (dialogue, ambience, SFX, music or silence — strongly recommended)
5. Shot structure (one continuous shot vs cuts — optional)

Always describe the audio. A prompt that says nothing about sound gets invented music.

A young chef in a crimson apron ladles steaming broth in a rainy night market, shallow depth of field, gentle camera push-in.
Single continuous shot, no scene cuts. Ambient rain, sizzling woks, soft murmur of customers. No music.

Five levers that make or break a prompt:

Negatives go inside the prompt — “Do not add captions” — there is no separate negative-prompt field.

Step 3 — Add references and keyframes

• Up to 10 images + up to 3s of video per generation
• Tag roles so the model knows what each file is for: FIRST_FRAME, LAST_FRAME, IMAGE_REF_0, VIDEO_REF_0 (numbered by upload order)
• Give each reference one job — style from the first image, character from the second — rather than leaving it to guess
• Keep video references short (≤3s, centered on the likeness — their audio is ignored)
• Loop trick: use the same image as both first and last frame → seamless loop

<FIRST_FRAME> <LAST_FRAME> A marble rolling fast on a chain reaction style track, single continuous shot, soft studio light. No music.

Tip: Attach everything up front. Adding a new reference mid-conversation can destabilize a scene that was holding.

Step 4 — Generate, then edit by talking

Hit generate. When a result is close:

• Pick the best take
• Type a short, surgical edit and keep the rest:

Make the phone invisible, keep everything else the same.

Change the lighting to be more dramatic, keep everything else the same.

Transport the violinist to the environment in IMAGE_REF_0, keep everything else the same.

Rules that save credits:

• One change per turn
• Re-anchor after ~4 edits
• Extend only appends to the end (10s → 40s)

Step 5 — Extend the scene

In Flow click Extend on the clip, or type:

Continue the video. Camera slowly pulls back as she keeps playing, warm evening light, same rain.

Each extension adds 10s and reads the last 10s of context — no jump cut if the prompt is continuous.

Step 6 — Draft cheap, upscale once

Generate 3–4 draft variations at 360p (compare side by side in Flow’s Draft Room). Vary one thing at a time. When one locks, download at 720p or upscale to 1080p/4K:

• Flow: Draft 360p → Download 720p → Upscale to 1080p (free for Plus/Pro/Ultra) or 4K (Ultra only, 50 credits)
• API: set resolution: "360p" for drafts, then re-render keeper at "720p" / "1080p" / "4K" or call upscaling separately

On the phone Flow app: draft at 360p on the go, upscale your favorite later on desktop.

In the Gemini API (code)

Text-to-video

import base64
from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-omni-1.1-flash",
    input="A marble rolling fast on a chain reaction style track, single continuous shot, soft studio light. No music.",
    response_format={"resolution": "720p"},
)

with open("marble.mp4", "wb") as f:
    f.write(base64.b64decode(interaction.output_video.data))

Image-to-video / first & last frame

from google import genai

client = genai.Client()

# Upload your images first via Files API, then tag them
interaction = client.interactions.create(
    model="gemini-omni-1.1-flash",
    input=[
        {"type": "text", "text": "<FIRST_FRAME> <LAST_FRAME> Smooth orbital camera move around the product, studio light, single shot."},
        # attach FIRST_FRAME and LAST_FRAME images here
    ],
    response_format={"resolution": "720p"},
)

Set the task explicitly when intent could be ambiguous:

interaction = client.interactions.create(
    model="gemini-omni-1.1-flash",
    input=[{"type": "text", "text": "Animate this product shot"}],
    # include image
    # video_config={"task": "image_to_video"}  # or text_to_video / edit / extend
)

Edit your own uploaded video (Files API)

import time, base64
from google import genai

client = genai.Client()

video_file = client.files.upload(file="Video.mp4")
while video_file.state == "PROCESSING":
    time.sleep(10)
    video_file = client.files.get(name=video_file.name)

if video_file.state == "FAILED":
    raise ValueError(video_file.state)

interaction = client.interactions.create(
    model="gemini-omni-1.1-flash",
    input=[
        {"type": "document", "uri": video_file.uri},
        {"type": "text", "text": "When the person touches the mirror, make it ripple like liquid and the arm turns to mirror material. Keep everything else the same."},
    ],
)

with open("edited.mp4", "wb") as f:
    f.write(base64.b64decode(interaction.output_video.data))

Large videos via URI delivery (avoids 4 MB payload limit)

interaction = client.interactions.create(
    model="gemini-omni-1.1-flash",
    input="A beautiful sunset over mountains, single continuous shot.",
    response_format={"type": "video", "delivery": "uri"}
)

video_uri = interaction.output_video.uri
# Poll until ACTIVE, then:
video_bytes = client.files.download(file=video_uri)
with open("output.mp4", "wb") as f:
    f.write(video_bytes)

Best-practice flags for speed: set background=false, store=false, stream=false for faster synchronous generation — but store=false means you cannot edit via previous_interaction_id next turn.

Prompting cheat sheet (from Google’s prompt guide)

• Force a single shot: “In a single continuous shot, no scene cuts”
• Silence: “No music, just room tone / realistic sound”
• Dialogue: she says: "Did your father go out on the boat too?"
• Text: name exact wording — otherwise random text appears; English renders reliably, other scripts vary
• Timing: “After 3 seconds, a woman enters” or timecode blocks every 2s
• Trigger: “When the person touches the mirror, …”
• Keep-else: always append “Keep everything else the same” on edits
• Aspect / resolution / duration are settings, not prompt text

Full guide: deepmind.google/models/gemini-omni/prompt-guide and ai.google.dev/gemini-api/docs/omni#prompt-guide

Practical workflow — what pros are actually doing

1. Storyboard at 360p — generate 3–4 cheap drafts varying one thing at a time (camera, light, or action)
2. Lock keyframes — pick the best draft, fix first/last frames if needed, set aspect
3. Iterate via conversation — 1–2 surgical edits with keep everything else the same
4. Extend if needed — grow from 10s → 20s → 40s with continuous prompts
5. Upscale once — render the keeper at 720p, then 1080p or 4K
6. If the brief needs native 4K, 21:9, or 15s+ single-take — hand that hero shot to Veo 3.1, Kling 3.0 or Sora 2 Pro instead

Customers already in production per Google: Adobe Firefly, Figma Weave, GMI Cloud, Runway, plus FLORA, Picsart, Cartwheel, Lovart, Artlist — all using Omni via the Agent Platform API.

Limitations to know before you sell it to a client

• 40s = stitched (4 × 10s). Joins are good (10s context) but test across all four extensions — not just the first.
• 4K = upscale. Check if your delivery truly needs native 4K vs upscaled.
• Two aspect ratios in API (16:9, 9:16). H3/Seedance offer 21:9, 4:3, 1:1, 3:4 too.
• Duration cap: 3–10s per generation. For a 25s continuous take, Sora 2 Pro wins.
• Audio: Omni outputs native audio (dialogue/SFX/music) — but audio-as-input is not currently processed; voice editing is restricted.
• Text: English reliable; other scripts vary.
• Complex motion / text accuracy still not perfect — Google’s own model card says so.
• Failed generations cost zero — you are not charged for blocked generations.

Official links — every URL you need

Wrapping up — should you switch?

If you already build with Veo 3.1, Seedance 2.0 or Kling, you don’t need to throw anything away. Omni 1.1’s advantage is iteration economics: draft cheap at 360p, edit by talking, extend to 40s with real context, upscale once. For a solo creator selling reels, ads or explainers, that loop is where the money is.

For the toughest shots — hard stunt physics, native 4K widescreen, 15s+ single take, or feeding real audio + lip-sync — keep Seedance 2.0 / Kling 3.0 / Sora 2 Pro in the toolkit and route by deliverable, not brand loyalty.

Start today: open Google Flow, switch the model picker to Gemini Omni 1.1 Flash, drop in one image and a one-sentence brief, generate a 360p draft, then edit it in conversation. In 15 minutes you will know more than any leaderboard can tell you.

Questions or stuck on Flow credits, prompting, or the Interactions API? Join the SPX WhatsApp community and drop your clip + prompt — we and other builders on the same release will help you dial it in.