What Happened
Google Cloud has made Gemini Omni 1.2 Flash generally available for video generation and editing, with API access, Provisioned Throughput, enterprise security capabilities and imperceptible SynthID watermarking. The model supports clips up to 15 seconds and lists a price of $0.07 per second for 720p video, 30% below version 1.1. [1]
The release adds more consistent reference-based generation, conversational editing and more natural speech. Drafts can be upscaled to 1080p or 4K, while audio references require allowlisting. Business access also includes Gemini for business and Google Vids; prompt- and product-photo-based video creation in Gemini chat is described as coming soon, not already available. [1]
Why It Matters to Businesses
The immediate opportunity is lower-cost production of short marketing clips, product videos and creative variants. Creative platforms including Artlist, Runway and Photoroom are already using the model, according to the announcement. That demonstrates integration activity, but does not establish the cost or quality a particular enterprise workflow will achieve. [1]
At the listed rate, one 15-second 720p generation costs $1.05. Ten thousand such generations would cost $10,500 before any additional charges. The useful business metric is cost per approved asset, not cost per generated second. Rejected outputs, revisions, storage, delivery and human review all affect that metric.
General availability makes the service a production integration option. It does not remove the need to validate output quality, capacity, latency and contractual requirements against the intended workload.
Kimbodo Engineering Perspective
We would treat video generation as an asynchronous media workload, not a blocking extension of a chatbot request. Submission, generation, review and publication have different failure modes and should have separate states. A browser disconnect should not lose a job, and a client retry should not automatically trigger another paid generation.
A managed API reduces the need to operate inference infrastructure, but shifts engineering effort toward orchestration, governance and supplier dependencies. We would isolate provider-specific calls behind an adapter while retaining prompts, asset metadata and workflow state in the application’s own data layer. That preserves operational control without assuming that different video models are interchangeable.
Provisioned Throughput is available, but the notes do not specify its pricing, capacity terms or service commitments. [1] We would evaluate it after measuring sustained demand and burst behavior, rather than assuming a reservation is cheaper or necessary.
How We Would Implement It
- Create a controlled submission API. Authenticate users, enforce tenant quotas, validate clip duration and reference assets, and reserve an estimated budget before accepting a job.
- Separate workflow state from execution. Store job records in a database, place work on a durable queue and keep reference media and outputs in private object storage. Track submitted, running, review-required, approved and failed states.
- Build a provider adapter. Centralize authentication, request construction, status retrieval and error handling. Maintain internal idempotency records; reconcile ambiguous provider responses before resubmitting work that may already be billable.
- Gate publication. Check technical properties and route outputs through policy and brand review. Require human approval where product accuracy, likeness, speech or advertising claims matter.
- Instrument unit economics. Attribute generated seconds, attempts, revisions and approvals to each tenant and campaign. Alert on spending spikes and retry loops; impose maximum attempts and per-job budgets.
- Evaluate before expanding. Test representative prompts and reference assets for consistency, speech quality and approval rate. Record model versions and configuration so changes can be assessed against a stable evaluation set.
Where the workflow supports it, we would approve a lower-resolution draft before upscaling. Higher-resolution output is supported, but its pricing is not established in the announcement; it should be budgeted separately. [1]
Risks, Costs and Security
Media inputs need explicit governance. Product images, unreleased campaigns, voices and faces may contain confidential information or require usage rights. Enforce tenant isolation, least-privilege service identities, short-lived asset access and defined retention periods. Verify provider data-handling, regional availability and contractual terms before sending sensitive material.
SynthID watermarking provides a provenance signal, not a substitute for consent, rights clearance or factual review. Enterprise security capabilities likewise do not automatically secure the surrounding application. [1]
The main operational risks are duplicate submissions, uncontrolled revision cycles, capacity constraints and publishing unreviewed outputs. Address these with budget enforcement, bounded retries, reconciliation and approval gates. A lower generation price improves economics only when the application controls how many generations it takes to produce a usable result.
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Infrastructure & MLOps practice, or Estimate My Infrastructure.
