Google has cut the published 720p video-output rate to $0.07 per second and added native 15-second clips and conversational editing. The open question is whether cheaper generation means fewer costly attempts to get a video a customer will approve.
Google's Gemini Omni 1.2 Flash adds native 15-second video creation, conversational editing and stronger reference following, according to the company's November 5 announcement. Google says the model is generally available through the Gemini API and Gemini for business. The published 720p video-output rate is $0.07 per second, 30% below Gemini Omni 1.1. Google Google Cloud
The list-price cut is straightforward. The economics of a completed video are not.
A model can generate a cheaper first draft and still cost more to use if it needs repeated attempts, introduces errors during editing or takes more human review time.
The relevant comparison is cost and hours per approved 720p clip. Investors cannot infer Google's margin from the new output rate.
The 30% saving covers billed video output, not production
At the published 720p rates, a 10-second output costs $1.00 under Omni 1.1 and $0.70 under Omni 1.2. That is a 30% cut in the published output charge for an identical 10-second 720p clip.
At the new rate, a 15-second output costs $1.05 before other charges.
These comparisons hold output duration and resolution constant. They do not show the total cost of a finished production asset.
Inputs, editing, higher resolution, upscaling, storage, moderation and human review can add costs, depending on the workflow and applicable billing.
Take an asset that Omni 1.1 produces acceptably on its first 10-second attempt. If Omni 1.2 needs an additional full-length generation, that second output adds another $0.70 and can erase the apparent price advantage.
This is an arithmetic sensitivity, not evidence that Omni 1.2 actually needs more attempts. The number of billed generations matters as much as the price of each one.
The opposite outcome is equally possible. If better reference following or conversational editing reduces failed attempts, the newer model can save more than its lower output rate suggests.
The announced prices establish neither outcome.
The cost measure must count failures without counting them as successes
A buyer comparing the two models needs a consistent measure of the cost of completed, approved work.
Cost per approved asset = (billed video output for drafts and edits + billed inputs and upscaling + review and other workflow costs) / number of approved assets
Each cost should appear once. Workflow charges such as storage and moderation belong in the total when incurred.
Rejected generations and retries go in the cost numerator. The denominator counts only approved clips.
This matters for advertising and commercial production, where an otherwise attractive video may be unusable because packaging text changes, a product's geometry drifts or a reference image is not preserved.
The production hypothesis is that longer native clips, conversational edits and closer reference following reduce costly reruns. Google's announcement does not measure that outcome.
A creative team might generate from a product photograph, revise a camera move or dialogue, and upscale the approved result. That workflow could reduce manual compositing and discarded renders if edits preserve the parts of the scene that should stay unchanged.
But extra edits also create costs and chances for error.
More editing turns can increase billed video output and review work. Token counts alone cannot be compared with a per-second video rate.
The native 15-second output limit adds another wrinkle. Longer assembled sequences can involve extensions and edits, each adding potential cost and quality risk. Continuity, lip-sync, text accuracy, reference identity and moderation outcomes therefore need evaluation across the full workflow, not just the first clip.
General availability does not establish production economics
Google's Cloud documentation describes model quotas and provisioned throughput. General availability, however, does not establish a particular buyer's queue times, concurrency, high-percentile latency or approval rates.
Those operational variables can matter as much as nominal output pricing when a team needs to deliver usable work on schedule.
Buyers also need to compare the applicable terms and entitlements across API, Cloud and business channels rather than assume every deployment has identical access and constraints.
The published rate makes an unchanged amount of 720p output cheaper. It does not establish that production adoption follows.
The most directly supported conclusion is therefore limited: Google has lowered the stated price for comparable video output.
A plausible upside is that editing and reference fidelity together also reduce wasted generations and review effort. A less favorable outcome is that revisions, quality failures or workflow delays absorb the price saving.
The announcement does not supply the acceptance-rate or production-cost evidence needed to choose between those outcomes.
A lower customer-facing output price also does not establish lower inference costs for Google, higher vendor margins or improved customer profitability.
A six-month test can separate lower prices from lower costs
A controlled six-month buyer test should compare billed and reviewed cost per approved 720p asset. External investors cannot observe that metric unless the buyer discloses its logs.
The buyer would run both models on the same prompt set, with equivalent reference inputs, duration, resolution and predefined approval standards.
For every usable final asset, it would record the billed output, failed generations, edits, upscaling, review hours and other relevant workflow costs. Approval rates, elapsed time and quality failures would be reported alongside the cost calculation.
The comparison would answer two separate questions: whether the output rate is lower, and whether the total cost of reaching an approved asset is lower.
CTOL's proposed 20% hurdle is an illustrative procurement choice, not a documented break-even point or a generally meaningful industry threshold.
A buyer could adopt that hurdle if it matches its economics. Other buyers may put more weight on turnaround time, quality or the predictability of their approval process.
CTOL's low-confidence call is that Omni 1.2 will lower total cost per approved 720p clip in a matched six-month workflow test. Billing, rejection and review logs would falsify that call if the cost does not fall.
The basis for the call is the lower published output rate. The low confidence reflects missing evidence on relative rejection rates, editing requirements and labor costs.
A test run over six months from the November 5 announcement would provide the relevant observation point, but its results would need to be disclosed before outside investors could use them as evidence.
Google has already made the first generation cheaper at the stated rate. What remains unproved is whether the full production process becomes cheaper.
For customers, the number that matters is the cost and time required to approve a usable video. For investors, the next step is evidence that such improvements influence commercial adoption or financial performance.
Until those results exist, a 30% reduction in billed video-output price should be treated as a pricing fact, not a demonstrated 30% improvement in production economics.