We benchmarked Gemini Omni Flash on Video Generation, and it took first place by 102 Elo points. Gemini Omni Flash ranks 1st overall on Design Arena’s Video Arena evaluation, 7 places higher than its highest-performing predecessor Veo 3 Fast. To do so, it surpassed the Seedance 2.0 model line, which has held the top spots on our leaderboard since its launch in February.
Gemini Omni Flash also sets a new Pareto frontier for preference vs price at approximately ~$0.10 per second of video compared to Seedance 2.0’s ~**$0.30 per second of video, **alongside being almost 2.5x faster.
How did Google Create the #1 Video Generation Model? #
To answer this question, we ran a case by case analysis on deployments of Gemini Omni Flash and observed how its optimizations have improved its performance across text-to-video generation tasks. This allows us to determine not only which optimizations are most effective, but also which error cases the model avoids.
We find that Gemini Omni Flash wins on the fundamentals of video generation, avoiding the physics and temporal-consistency failures that sink other models and following prompts down to specific details. Compared to other top models that win on raw cinematic polish, Omni Flash focuses on predicting and responding to a user’s intent, maintaining a creative consistency that other models lack. The model is not flawless, but with over 5,000 tournaments, it has maintained a 1st place ranking in Video Arena.
Model Behavior #1: Phenomenal Physics #
Much of Gemini Omni Flash’s high performance in video generation can be attributed to improved physics. Objects move with realistic weight and scenes stay coherent, avoiding the unrealistic breaks that other models often fall into.
The model excels at retaining temporal consistency, the ability to avoid randomly generating new objects or repeating scenes. Generations flagged for temporal breaks win 65.4% of tournaments, while clean generations win 72.4%,** a 7-percentage-point swing.**
The same holds for its ability to generate natural, fluid motions. Generations flagged for jerky or unnatural movement win 65.9% of tournaments, while smooth generations win 71.1% (baseline 66.1%). Gemini Omni Flash generations are flagged less often than its competitors, leading to increased user preference.
Model Behavior #2: Detail-Driven Dominance #
One of Gemini Omni Flash’s strongest aspects is its ability to respond and adapt to user prompts. It shows prompt adherence on 81% of generations compared to the 63% average of all models on Design Arena. When a prompt asks for a specific object or action, Omni Flash delivers, accomplishing every task the user requested in a prompt.
Its detail-oriented performance carries over to creating legible text, one of the most persistent failure cases in video generation. Where other models may produce smeared, shifting letterforms, Omni Flash renders legible on-screen text, letting it excel at the educational, film, and marketing-focused prompts that make up a plurality of user requests.
Gemini Omni Flash exhibits strong animation capabilities as well, excelling in photorealistic footage to stylized 3D animation without compromising on prompt details.
Prompt adherence gives Gemini Omni Flash its characteristic strengths in animation and product-focused videos since these subcategories benefit the most from a video model’s ability to answer a prompt to a user’s satisfaction.
Model Behavior #3: Lacking Landscapes and Extreme Emotions #
However, Gemini Omni Flash isn’t a perfect model. It performs significantly worse when generating videos of landscapes, losing 14.5 percentage points in win rate on the 3.5% of Landscape prompts. This is often due to a lack of realism compared to competing models, with Omni Flash creating scenes that look more animated than picturesque.
Additionally, when explicitly prompted for emotions, Gemini Omni Flash tends to exaggerate emotions which leads to videos feeling “uncanny.” While not prevalent enough to significantly impact win rates, it may be a signal of Omni Flash over-optimizing for prompt adherence instead of realism.
What this means for model selection #
Gemini Omni Flash wins for a reason that should reshape how people pick video models: not spectacle or visual flair, but reliability and fidelity. For designers and directors looking for a video generation model, Gemini Omni Flash brings noticeable improvements in realizing creative insights and bringing visions to life. Launches like this are a reminder of how fast the video generation frontier is moving, with new models outperforming older models at a third of the price and 2.5x the speed.
We will continue monitoring Gemini Omni Flash's performance and how it compares to other models. Congratulations to the @GoogleDeepMind team on the launch.
Direct your own videos now and see if you prefer Omni Flash's generations on DesignArena.ai.
Written by @anmayg_ and the @Intelligence_ai team.