Quick Answer
Gemini Omni Flash entered the Gemini API in public preview on June 30, 2026 under the model ID gemini-omni-flash-preview. Google describes generation and conversational editing from combinations of text, image, and video inputs.
Design the workflow around current preview facts:
- output is currently 10 seconds;
- launch price is $0.10 per second of output;
- audio-reference upload is not supported in the API;
- scene extension is not supported in the API;
- short video references can be accepted by the schema but Google says they are not processed correctly at launch;
- character consistency can weaken across scene changes or pans;
- the Interactions API can preserve context for up to three sequential edits in Google’s launch workflow;
- SynthID watermarking is applied.
Do not freeze the preview model ID or limitations in a long-lived integration without a live documentation check. Use the Gemini Omni Flash model guide for the current entity record and the product vs API guide for access-surface selection.
Source check: Google’s June 30 developer launch, July 1 recap, DeepMind model page, and developer documentation were checked July 30, 2026.
Step 1: Confirm the Developer Surface
Google lists Gemini Omni Flash on:
- Google AI Studio;
- Gemini API;
- Gemini Enterprise Agent Platform.
This guide is about the Gemini API preview. Gemini app and Google Flow access are product workflows with different UI, subscription, geography, and entitlement boundaries.
Before coding, verify:
- the model is available in the intended region and project;
- the project has the required billing and quota;
- the current model ID;
- current pricing and request limits;
- preview terms and lifecycle;
- data-use, retention, safety, and enterprise controls;
- supported SDK or REST schema.
Google product or API availability does not establish that the model is available in Flowith. Verify Flowith separately.
Step 2: Define One Accepted Clip
Write the acceptance criteria before prompting:
| Field | Record |
|---|---|
| Purpose | What the clip must communicate |
| Duration | Current output target within the documented 10 seconds |
| Inputs | Text, image, and/or video reference |
| Subject | Appearance, wardrobe, object, or brand traits |
| Scene | Environment and time |
| Action | One clear sequence that fits the duration |
| Camera | Framing, movement, and transition |
| Text/graphics | Exact approved copy and timing |
| Continuity | Traits that must remain stable |
| Safety/rights | Permission, trademark, person, location, and disclosure checks |
| Acceptance | Visual, factual, motion, text, and technical pass/fail rules |
A short output benefits from one coherent action. Compressing several scenes into ten seconds increases continuity and review risk.
Step 3: Prepare Inputs
Google describes combined text, image, and video input. Use the smallest input set that controls the intended result.
Text
Specify:
- subject and environment;
- action order;
- camera behavior;
- visual style through descriptive traits;
- exact on-screen text;
- prohibited changes;
- reviewable output requirements.
Avoid imitating a living artist or copying protected characters and brand assets without permission.
Image
Use an image when composition, product shape, character appearance, color, or setting needs a visual anchor. Confirm you own or are allowed to use it.
Video
Treat video-reference support conservatively. Google’s launch post says references up to three seconds are accepted by the API schema but are not processed correctly by the model at launch. A schema-valid request is not proof of useful model behavior.
Do not build a required production workflow on that reference path until the live documentation says the limitation is resolved and your test passes.
Audio
Google says audio-reference upload is not supported in the API for this model at launch. Do not silently strip audio and imply it influenced the result.
Step 4: Build the First Generation
Use the current Gemini Omni documentation for the exact request schema and SDK. Preview APIs change, so copying an old code sample can create a false implementation contract.
For the first request:
- use the documented preview model ID;
- send the minimum verified inputs;
- request one ten-second-or-shorter action;
- log model ID, date, region, request settings, and input references;
- retain the provider response and error metadata allowed by your data policy;
- validate the output before any follow-up edit.
Do not treat an HTTP success as an accepted clip.
Step 5: Review the First Output
Review frame by frame:
- subject and object consistency;
- anatomy and geometry;
- camera path;
- physical continuity;
- exact text and graphic timing;
- brand and product details;
- factual claims;
- unwanted additions;
- safety and rights;
- technical format;
- disclosure and provenance requirements.
Google says character consistency has limitations during scene changes and panning. Include those movements in the acceptance test if they are important.
Step 6: Use Conversational Editing Deliberately
Google positions Omni Flash for natural-language video editing. Its launch workflow says the Interactions API can preserve session history and context for up to three sequential edits.
Use one correction per turn:
- preserve the approved subject and composition;
- name the exact defect;
- state the desired change;
- state what must remain unchanged;
- rerun the full acceptance review.
Example edit brief:
Keep the subject, product shape, framing, and approved on-screen text unchanged. Slow the camera pan and move the product reveal to the final two seconds. Do not add new logos, people, or props.
Do not assume a later edit preserves every earlier detail. Save accepted versions and compare them.
Step 7: Budget Accepted Output
Google’s launch price was $0.10 per second of output. At the documented ten-second length, one generated clip implies $1.00 of output-generation cost before taxes or other workflow costs.
Use:
accepted clip cost = generated seconds × live rate × attempts per accepted clip + storage + review + post-production
Example scenario:
| Input | Value |
|---|---|
| Output per attempt | 10 seconds |
| Launch rate | $0.10/second |
| Attempts per accepted clip | 4 |
| Generation subtotal | $4.00 |
| Human review and post-production | Add measured cost |
This is an illustration using the launch rate, not a current quote. Verify live pricing and how failed or filtered requests are treated.
Step 8: Design Around Current Limitations
Ten-Second Output
Plan a self-contained clip. Do not promise longer duration based on Google’s statement that it is coming later.
No API Scene Extension
Do not make scene extension a required step. If several clips must connect, create a shot plan and verify continuity through editing.
No Audio Reference
Plan audio separately and verify rights, synchronization, loudness, captions, and disclosure in post-production.
Video Reference Limitation
Disable or label the path until live documentation and your test show it works correctly.
Character Consistency
Reduce scene changes and large pans when identity consistency is critical. Use stable reference inputs and grade every frame.
Step 9: Add Safety and Provenance Controls
Google says Gemini Omni uses SynthID watermarking. Watermarking is one provenance signal, not a substitute for:
- user-visible disclosure where appropriate;
- consent for recognizable people;
- rights to input images, footage, audio, logos, and text;
- organizational policy review;
- output moderation;
- audit logs;
- platform-specific labeling.
Do not use the model to deceive viewers, impersonate people, or create misleading evidence.
Step 10: Create a Preview Migration Plan
Public preview can change:
- model ID;
- request schema;
- limits;
- price;
- regions;
- quality;
- safety behavior;
- deprecation timing.
Before production:
- isolate the model ID in configuration;
- log the actual returned model/version signal when available;
- pin tested SDK dependencies;
- validate responses;
- set cost and retry limits;
- create a fallback path;
- monitor official release notes;
- rerun the acceptance set before switching versions.
FAQ
What is the Gemini Omni Flash model ID?
Google’s launch names gemini-omni-flash-preview. Verify the live documentation because preview identifiers can change.
How much does the API cost?
Google published $0.10 per output second at launch. Budget retries, rejected outputs, review, storage, and post-production, and verify the current rate.
Can I generate videos longer than ten seconds?
Google documents ten-second output at launch. Longer duration was described as future work, not current availability.
Can I upload an audio reference?
Google says audio-reference upload is not supported in the API for this model at launch.
Can I extend a scene?
Google says scene extension is not supported in the API at launch. Design a multi-shot editing workflow instead of depending on it.
Bottom Line
Build the preview workflow around a ten-second accepted clip, not around a feature roadmap. Keep product access, API preview, model ID, price, region, and Flowith availability separate. Use strict input permissions, one-change editing turns, frame-level review, a cost ceiling, and a migration plan.
For the surface decision, continue to Gemini Omni Flash product vs API access. For the source-backed model summary, use the Gemini Omni Flash model page.