15 Seconds, 7 Cents per Second: Gemini Omni 1.2 Flash Changes Both the Video Leaderboard and the Price Tag

On November 5, Arena.ai published its latest video model leaderboard. Google's Gemini Omni 1.2 Flash ranked first in Text-to-Video and Image-to-Video, and third in Video Edit.

The scores are specific:

Text-to-Video: 1541 points, 27 points higher than Gemini Omni 1.1 Flash's 1514 points. Google's Omni series therefore took the top three spots in text-to-video.
Image-to-Video: 1502 points, 39 points higher than the previous Gemini Omni Flash's 1463 points. The previous version ranked 8th.
Video Edit: 1395 points, ranking 3rd, 20 points higher than Omni Flash's 1375 points. The previous version ranked 5th.

Google AI's release post lays out the changes clearly: cheaper; the upper limit for video generation and editing extended to 15 seconds; scenes more stable over time; upgraded character speech; fewer visual glitches such as jump cuts; and greater fidelity to uploaded reference material.

The official demo takes a single style frame and generates a 15-second macro scientific explainer of a honeycomb with 4 shots. The video is here:

Available entry points are also provided: import scenes and do multi-round editing in Flow; generate 15-second clips for free in Google Vids; create 15-second videos in the Gemini App; and developers go through the API. 720p video costs $0.07 per second, 30% lower than the previous generation. 15 seconds is about $1.04. 1080p upscaling is $0.104 per second. Calls on the previous generation 1.1 at $0.10 per second will be billed at 1.2 rates starting December 3.

There is also a screenshot of the text-to-video leaderboard in Arena's replies:

With the price and ranking changing at the same time, it is easy to look only at the first line. Commenter NotaDEV said that having cheap and first appear in the same release is indeed rare. What is more worth calculating is batch cost. Anand Butani's reply was direct: 720p at $0.07 per second, about $1.05 for 15 seconds; this pricing is aimed at marketing teams batch-testing ad variants, and filmmakers are not necessarily the target. The 30% reduction is amplified when there are many calls. For people doing product shots, Jeff Bruchado warned that reference fidelity matters more than the extra few seconds; if the logo deforms midway, the whole video is ruined.

The replies under the leaderboard are not so tidy. Some asked whether they could run it on their own hardware, or whether they could only use Flow, Vids, and the API; some asked when Google would open-source these models. Toby VirtualFilmer questioned why a model can rank first on the leaderboard as soon as it is released, and whether freshness is a ranking factor. FF said that Arena has not made clear what exactly it tests; he tested with reference images, and Seedance 2.5 completely beat all previous models. Hanchi Sun said that without Seedance 2.5 at the top, it is just wrong. These are the doubts in the replies; only by looking at them together with the leaderboard scores does it feel like a complete launch scene.

A complaint from commenter Fang Buxiu fits here: the video generation leaderboard changes leaders faster than turning pages; image quality is no longer the only variable, and how long can be generated at once and how much it costs per second have become more practical thresholds.

For ordinary users, 15 seconds and free entry points are more direct than leaderboard rankings. For developers, the API price list and the December 3 rate switch are more important. The leaderboard's first place will keep changing. Next time you look at a video model, first turn to the duration limit and price list, then go back and watch the demo.

Arena.ai original tweet: https://x.com/arena/status/2118420238756921648

Google AI original tweet: https://x.com/GoogleAI/status/2118389961426874879

Arena leaderboard: https://arena.ai/leaderboard/video/image-to-video