Most AI video products expose separate models for generation, reference control, and editing. Gemini Omni Flash tries to make those modes one conversation. Google introduced it on June 30, 2026 as a fast multimodal video model that can interpret text, images, and supported video, then create or revise a clip with audio.
The product name invites an important qualification. “Omni” does not mean every possible input and output. The public-preview API supports a defined set of tasks and media limits. Within those boundaries, however, it offers a notably broad workflow: start from a prompt, a still image, several references, or a short clip; generate video; then request another edit using the prior interaction.
One model, four task modes
Google documents four task types: text-to-video, image-to-video, reference-to-video, and edit. Text-to-video creates a scene from written direction. Image-to-video animates a supplied frame. Reference-to-video uses supplied images as visual evidence for elements or appearance. Edit transforms supported video according to instructions.
Using one model across these modes can reduce handoffs. A developer does not need to translate an edit request into the prompt dialect of a separate system. The conversation preserves a semantic thread: generate a shot, change an object, alter the camera treatment, or refine the motion.
Task declarations still matter. The API needs to know which operation is intended, and each mode has its own valid inputs. Production applications should validate file type, size, duration, aspect ratio, and task compatibility before sending a request.
Output specification and context
The current model card lists gemini-omni-flash-preview with a 1,048,576-token input context. It generates video rather than text. Google documents clips from 3 to 10 seconds at 720p and 24 frames per second. Those limits make it a shot generator or editor, not a complete long-form renderer.
The large context is useful for understanding multimodal directions and references, but it does not extend the output duration. A campaign may include many instructions and images while each rendered result remains short. Teams should split a longer concept into shots and preserve a structured storyboard outside the model interaction.
Every generated video includes SynthID, Google’s imperceptible watermark for AI-generated content. That assists provenance but does not replace editorial disclosure, rights review, or a record of which sources and prompts shaped the result.
Audio is part of the generated clip
Gemini Omni Flash attempts to generate audio with video. Prompts can describe ambience, sound effects, and spoken content. This can make previews feel more complete than silent animation and helps a creative team judge rhythm before post-production.
Audio is not guaranteed to match exact production requirements. Dialogue may need timing, pronunciation, language, voice, loudness, or licensing controls that a generated soundtrack does not provide. Google’s current limitations say audio references and voice editing are unsupported. The model also does not provide a negative-prompt parameter, so unwanted audio must be handled through clearer instructions, another iteration, or downstream replacement.
Treat generated sound as a creative output to review, not a locked master. Keep scripts and approved voice assets separate so they can be replaced deterministically.
Reference images and conversational editing
Reference inputs let a prompt point to supplied visual elements instead of describing everything from scratch. The API guide supports multiple image references with tags, which can be helpful for a product, wardrobe, setting, or art direction. Exact consistency should still be tested across the angles and motion a project needs.
Conversational editing uses an earlier interaction as context. A developer can ask for a revision without reassembling every previous input. This is powerful for exploration because feedback can be expressed in normal production language.
Interaction storage has operational consequences. If an application sets storage off, the prior interaction cannot later be addressed through previous_interaction_id. Teams need a retention policy that balances edit continuity with privacy. Store source provenance and approved outputs in their own asset system regardless of conversation retention.
Current preview limitations
Google’s documentation is unusually explicit about gaps. Audio references are not supported. Short video references may be accepted by the schema, but the current guide warns that they are not processed correctly. Multiple-video reasoning is unavailable. Extension and interpolation are not supported, and a generated voice cannot be selectively edited.
Video editing input is limited to short material; the model card describes supported video input up to 10 seconds, while the task guide notes tighter behavior for some reference uses. Developers should follow the latest task-specific documentation and reject unsupported requests locally.
System instructions, temperature, top_p, stop sequences, and negative prompts are also unsupported for this model. Provisioned throughput is not available, and YouTube URLs cannot be used as sources. Google says English is fully evaluated while behavior in other languages may vary. Region availability must be checked for the deployment.
These are public-preview conditions, not necessarily permanent product boundaries. Pin the model ID, monitor release notes, and run regression clips whenever the API changes.
Build that regression set before launch. Reuse a small collection of licensed prompts and references covering people, products, camera motion, dialogue, typography, and low-light scenes. On every model update, compare prompt adherence, subject continuity, audio usefulness, safety behavior, latency, and failure codes. Archive selected outputs with their request metadata. A visually impressive new version can still regress on the one product shot a production workflow depends on.
Price and throughput planning
The current Gemini API pricing page lists no free tier for Omni Flash. Paid input is $1.50 per million text, image, video, or audio tokens. Generated 720p video is priced at $17.50 per million output tokens, with Google’s published conversion of 5,792 tokens per second. That works out to about $0.10 per generated second before input, retries, storage, and finishing.
An eight-second output is therefore roughly $0.81 in video output charges at that rate. This is list-price arithmetic, not a promise about total job cost. Failed attempts, alternative takes, and edit rounds can dominate. Track accepted seconds per concept, not only generated seconds.
Preview capacity can fluctuate. Build asynchronous job handling, timeouts, bounded retries, and spend limits. Cache or reuse approved references where policy permits, and generate low-risk variations before committing to a large campaign batch.
Where Omni Flash fits best
The model is well suited to fast shot ideation, social clips, motion studies, product concepts, and short edits where text, imagery, and existing video need to be interpreted together. It is less suitable for long continuous scenes, precise voice replacement, frame-exact interpolation, or guaranteed multi-shot character continuity.
Use human review for anatomy, logos, text in frame, cultural context, and rights. Verify that reference assets may legally be used for generation. An API accepting an image does not establish permission to reproduce its subject or style.
Finishing an exported clip with Medux
Gemini Omni Flash and Medux serve different stages. Omni performs generative creation or transformation. Medux can apply defined media operations to the exported file through separately authorized MCP workflows. There is no claim that Google’s preview model natively calls Medux.
If generated sound is not usable, the Claude MCP audio-track tutorial shows how to attach a chosen track as a separate job. If branding needs to be exact, the Codex MCP logo tutorial covers a deterministic overlay. Several approved Omni shots can then be assembled using the Claude MCP video-merge workflow.
Preserve the boundary in the production record. Store the Omni model ID, prompt, references, interaction ID, generation date, and SynthID-bearing output. For each Medux operation, keep the input hash, parameters, task ID, status, resulting file, and reviewer decision. Do not describe the finished file as raw model output after it has been altered.
A robust sequence is generate, review, lock the selected clip, apply one approved finishing operation at a time, validate duration and streams, then publish. Gemini Omni Flash makes the creative front end unusually flexible. Explicit asset and job records make that flexibility compatible with a repeatable media pipeline.