Google cuts AI video prices by 30%: Gemini Omni 1.2 Flash makes 15-second clips for about a dollar
Google DeepMind has launched Gemini Omni 1.2 Flash, which extends videos to 15 seconds and cuts the per-second price by 30% compared with its previous version. For creators and small businesses in Latin America, questions remain about the final cost, real-world quality and availability by country.
Photo: Alan Alves on Unsplash
A 15-second promotional video featuring a store's product, a model's face and the brand's colors, for around a dollar. That is what Google DeepMind is offering with Gemini Omni 1.2 Flash, its video-generating artificial intelligence model. The company announced on November 5, 2026, that the model is now generally available in the Gemini API and on Google Cloud. The API is the technical channel through which other programs and services connect to the model. With this version, clips go from 10 to 15 seconds and, according to Google, the price drops 30% compared with Omni 1.1, to US$0.07 per second at 720p resolution.
The figure sounds attractive, but it is worth reading carefully. The 30% reduction is measured against the model's own previous version, not against rivals such as Veo, Sora, Runway or Kling. Omni 1.1 cost about US$0.10 per second. That figure comes from applying the 30% and from a DEV Community article citing Google's pricing page, and it has not been verified in an official source.
The website Unite.ai ran the numbers using the token rate; tokens are the small units these systems use to measure what they receive and what they produce. The paid API charges US$12 per million tokens of generated video, at a rate of 5,792 tokens per second. That comes to about US$0.70 for a 10-second clip and US$1.04 for a 15-second one. At the US$0.07-per-second rate, the 15-second clip comes to US$1.05; the difference is due to rounding. Input to the model (text, images or video) costs US$1.50 per million tokens, and there is no free tier. According to the same outlet, on December 3 requests directed to Omni 1.1 will automatically switch to version 1.2, at the new rate.
A secondary blog compared per-second prices at 720p: Veo 3.1 Fast at US$0.10 with 8-second clips, Kling 3.0 Pro via Runway at US$0.17 with audio, and Sora 2 Pro at US$0.30. Those figures are not confirmed in official sources, so they serve as a rough reference rather than a definitive table.
What it promises for brands and online stores
The model is aimed at advertising and commercial content. It supports combined references: images of a character, a product or a visual style, motion clips of up to 3 seconds, and audio tracks. Google says faces, clothing, packaging typography and brand colors stay consistent even when the camera angle changes. It demonstrates this with a 15-second virtual try-on. That consistency matters because a brand cannot publish an ad in which the logo changes from one shot to the next.
This is where the first discrepancy appears. Segmind, a firm that tested the model, says that in its access the reference mode only reached 10 seconds. Unite.ai indicates that the documentation uses 10 seconds by default. Something similar happens with conversational editing, which according to Google lets users request, turn by turn, changes to the product, background, lighting, weather or screen format across the entire clip. In Segmind's tests, the adjustment feature responded that it was not yet available. Google recommends making drafts in 360p or 720p and taking the final version to 1080p or 4K, but the cost of that last step does not appear in the published information.
On quality, the only numerical comparisons come from DeepMind itself: its raters preferred version 1.2 over 1.1 on visual consistency by 62% to 38%. Google did not publish scores for reference fidelity, lip sync or physical realism. Segmind spent US$16.17 on 20 requests. Responses took between 28 and 136 seconds, usually between 40 and 90. The firm called it the most instruction-faithful model it had tested, although it missed about one in four instructions. It also does not consider it suitable for realistic cinematic scenes with people.
Every video carries SynthID, an invisible watermark that identifies the content as AI-generated. One analysis warns that this watermark does not prove a scene is real, nor does it guarantee it will always be detected.
The open question in Latin America
For a creator or small business in the region, the main unknown is practical. There is no official information on availability by country, local pricing, billing in local currency, VAT or digital taxes. The price is published in dollars and, with no free tier, an active billing account is required. One source notes that editing or extending uploaded videos is not available in the European Economic Area, Switzerland and the United Kingdom, without mentioning restrictions for Latin America.
The launch drew attention on Hacker News, the tech forum where the thread garnered 189 points and 146 comments. Complaints about access to, naming of and pricing of Google's AI products dominated there, more than analysis of the model.
There are still things to watch for: the official pricing table by resolution, the cost of upscaling to 4K, independent quality tests, and performance with Spanish-language dialogue and Latin American faces. The terms for commercial use are also missing. Until then, a dollar per clip is a good starting point for running the numbers, but not the final price of a campaign.