Gemini Omni 1.2 Flash cuts video prices. What does a 15-second clip cost?

Google released Gemini Omni 1.2 Flash on November 5, extending generated videos from 10 to 15 seconds and advertising 720p output at US$0.07 per second, 30% below its predecessor. At that rounded rate, a 15-second clip comes to about US$1.05. The model is generally available on Google Cloud and accessible through the Gemini API.

For a small retailer making product clips or a team preparing social ads, the practical change is room for a longer sequence at a lower output price. Whether that reduces the cost of a finished video depends on how many generations survive review and how much editing they need.

The headline price is an estimate, not the whole bill

Google Cloud’s pricing schedule bills Omni 1.2 Flash video output at US$12 per million tokens. A second of 720p output uses 5,792 tokens, including audio. Multiplication gives US$0.069504 per second, or US$1.04256 for 15 seconds, before other charges. That explains the rounded seven-cent figure.

Input media and text, text output and reasoning have separate token charges. Upscaling also has its own output-token accounting. A low-resolution draft followed by a selected upscale may be useful, but the advertised 720p figure should not be treated as an all-inclusive 4K production price.

The other variable is acceptance. If a hypothetical 15-second deliverable takes four full generations, those attempts alone would total about US$4.20 at the rounded output rate, before edits, inputs or review time. This is an example of budgeting arithmetic, not a measured Omni success rate.

Editing keeps the conversation attached to the clip

The developer documentation supports durations from three to 15 seconds, with 720p as the default output resolution. It also lists 360p, 1080p and 4K options and landscape or portrait aspect ratios. Generation can take more than a minute; clip duration is not a promise of processing speed.

The API distinguishes stateless generation from stored interactions. Keeping the interaction state allows a later request to reference its ID for multi-turn video editing. Stored and asynchronous interactions are retained for up to seven days, an operational detail for anyone building a revision workflow around the service.

Google’s launch examples show follow-up edits changing a scene’s background while retaining its characters and framing. The company also says the model follows image, short motion-video and audio references more closely. These demonstrations illustrate the intended workflow; they do not establish how reliably it will preserve every product, face or written label in an unfamiliar brief.

Editorial visualization

For production work, the relevant test is a controlled revision: keep the subject and camera request fixed, change one background detail, and inspect whether anything else moved. A successful background swap is less useful if the product’s shape or packaging changes with it.

A free trial route, with a different allowance

Google says Vids users with a Google account or Workspace subscription can generate up to three 15-second videos per month at no cost. That offers a way to try the model before integrating an API. It is a specific monthly product allowance, not a promise of unlimited free API generation.

Three trials can reveal obvious mismatches with a brief, but they are too few to estimate a dependable production acceptance rate. Teams evaluating a catalog or campaign need representative subjects, fixed review criteria and a record of rejected takes as well as the attractive ones.

What the release does not yet prove

Google acknowledges in its announcement that fast sports action, rapid script changes and complex physics still need improvement. Its claims of stronger consistency and more natural speech should be treated as vendor claims. UpThat has not independently benchmarked this release or measured its retry rate.

Editorial visualization

The DeepMind model card lists text, image, audio and video inputs, with video and audio output. It says generated content carries SynthID watermarks and that the ability to change people’s speech is restricted for now. Those safeguards and limits are relevant to deployment, but they are not evidence that a particular generated scene is accurate.

Before approving a product video, review the subject’s geometry, colors and labels across the whole clip, then check speech and lip-sync separately. Record the total generation and editing spend per accepted shot. That is the comparison that will show whether the cheaper second has become a cheaper finished video.