Gemini Omni 1.2 Flash: Longer Clips, Lower Video Costs

Gemini Omni 1.2 Flash makes a specific video-production decision more interesting: whether one generated shot can carry a complete short sequence without needing another generation and a manual join. Google’s November 5 release adds native clips up to 15 seconds and reduces video-output pricing. For creative teams, the useful question is whether those changes reduce the cost of an accepted clip, including revisions and review. Google’s announcement

That distinction matters when the subject is a real product. A longer shot is valuable only if its label stays readable, its shape stays recognisable and the requested action finishes properly. A lower generation price helps, but it cannot compensate for footage that changes what you are selling.

Google positions the release around longer native generation, conversational editing and closer adherence to references. Those are provider claims, not a controlled comparison of commercial production work. The reason to try them is straightforward: some creative ideas are awkward when divided into shorter independent shots. Release overview

Consider a proposed product reveal. The camera approaches a closed presentation box, the lid opens, and the shot settles on the item. If the reveal finishes before the product can be seen clearly, the asset may be unusable even when individual frames look polished. More duration provides room for the action and the final hold within the same generated sequence.

It also creates more opportunity for things to drift. A logo can start correctly and become distorted during a turn. An object can change proportions as the camera moves. An attractive final frame does not establish that the frames preceding it are acceptable.

The headline maximum does not describe every video operation. A first-and-last-frame interpolation asks the model to connect two supplied states. Extension asks it to continue an existing sequence. Native generation starts a new shot. Those jobs create different continuity problems, so they should be tested separately.

Limits are from the Gemini model card and November 5 changelog, checked November 10. These are operation limits, not guarantees that every requested clip will be usable.

For a repeatable workflow, make the operation an explicit input alongside the prompt. A template that says only “produce a 15-second video” leaves a caller to guess whether it should create, interpolate or extend. That uncertainty becomes harder to diagnose when several tools share the same production queue.

Keep continuity checks close to the operation. For interpolation, compare the supplied endpoints with the resulting start and finish. For an extension, review a short window on either side of the join at normal playback speed. For generation, review whether the final action lands before the clip ends. Do not borrow a duration setting from an older model’s example without checking the new operation’s contract.

The paid Gemini API lists input at $1.50 per million tokens, text output at $9 and video output at $12. Video consumption varies by resolution. At 720p, the documented rate is 5,792 tokens per second, making a 15-second output calculation 15 × 5,792 × $12 / 1,000,000 = $1.04256. Gemini pricing

The practical budgeting error is to attach the lowest advertised figure to every deliverable. A team may explore at a lower resolution, select a candidate and then produce a higher-resolution result. The discarded explorations still belong in the job’s cost. So do any charged revisions and additional inputs.

Save the model ID, output settings and returned usage for each request. Aggregate at the creative-job level rather than reporting only the final call. The finished-second cost framework explains why generated duration and usable duration are different denominators. This release changes the price inputs; it does not remove that accounting distinction.

Google describes the price reduction as 30%. Comparing the listed old and new video-token prices, $17.50 and $12 per million, gives approximately 31.4%. The two statements use different precision; neither is a measured reduction in the cost of a completed campaign. Current and previous video rates

Use a generation allowance that matches the brief. For illustration, four 15-second 720p outputs would have a combined video-output component of $4.17024. If only one is accepted, that whole amount belongs to the accepted clip. It would be misleading to report $1.04 as the job’s total cost simply because that was the output component of the last request.

The example is deliberately limited. It assumes four equally long outputs, no other charges and one accepted result. It says nothing about how frequently Omni succeeds. Its purpose is to make the budget visible before a producer starts experimenting.

A useful production sheet has separate columns for model charges, editor time and approver time. Add a reason when a candidate fails: wrong product detail, unfinished motion, poor composition, incorrect audio or a changed brief. Otherwise, a change in creative direction may be blamed on model quality, while repeated model errors disappear inside a general editing budget.

For a product video, review can start with a small list of protected details: product name, package shape, visible quantity, principal colour and logo placement. These are examples of brief-specific criteria, not a universal score for image fidelity. A fabric campaign might care about texture and drape; a software demonstration might care about interface text and action order.

Create a reference sheet with an approved image and a short written description of each detail. Keep subjective preferences separate from factual requirements. “Feels more premium” invites discussion; “the package has the correct printed name” can be checked directly.

The review should include both still frames and playback. Still frames reveal label drift and malformed details. Playback reveals flicker, discontinuity and movements that look unnatural despite acceptable individual frames. If the deliverable will include sound, assign a separate check for whether the audio fits the approved content and timing.

For an initial comparison, reuse the same brief, references and acceptance criteria across versions. Record the settings and attempt budget. Randomise the order in which a reviewer sees candidates where practical, so the model name does not become a quality cue. The broader identity-blind evaluation guide discusses that principle.

Conversational editing is appealing because a producer can ask for a specific correction without rewriting the complete creative instruction. The important operational question is what happens to details that were already correct.

Suppose the first candidate has the right packaging but the camera stops too low. A request to raise the camera should trigger two checks: did the requested framing change, and did the packaging stay correct? Passing the first check is not sufficient. An edit that fixes composition while damaging the product requires another decision, not automatic approval.

Retain the original candidate and each selected intermediate result. Record the instruction that produced the next version. If the third revision becomes worse, the producer should be able to return to the last acceptable state without reconstructing it from memory.

Use small, explicit changes during evaluation. Asking for a new background, new lighting, a different camera move and corrected text in one turn makes it difficult to identify which request caused a regression. That does not mean a production system can never combine edits; it means a test should be interpretable before the workflow is automated.

Google’s November 5 notice says requests to gemini-omni-1.1-flash will route automatically to gemini-omni-1.2-flash from December 3, 2026, using the newer rates. An unchanged string will therefore stop being evidence of unchanged model behaviour. Gemini API changelog

Inventory where the old ID appears: application configuration, scheduled generation jobs, prompt examples, cost estimates and test fixtures. A model setting buried in a saved campaign template is easy to miss if the review examines only source code.

Run representative briefs against the explicit new ID before the routing date. Include at least one ordinary job and the awkward cases the team already knows: a particular logo, a tight crop, a reflective surface or a difficult continuation. Keep the evaluation scope manageable; the purpose is to detect meaningful regressions, not collect a gallery of attractive samples.

Review the downstream assumptions too. A system that assumes every successful request returns one particular duration may handle a longer result badly. File processing, storage, editorial review and delivery specifications each need to agree about what the job requested. A successful API response alone does not show that the finished workflow is correct.

The best first trial is a job the team already understands. Pick a short sequence with a clear acceptance standard and enough recurring value to justify the setup. Preserve the existing production route while learning whether the new model improves that specific job.

This proposed evaluation is designed to make the production decision testable. Its value is that the decision can be made from observed outputs and job costs. It avoids choosing a model solely because its headline duration is longer or its advertised second is cheaper.