Gemini Omni Flash
Gemini Omni Flash
One-line summary: Google's May 2026 conversational video generate-and-edit model; as of 2026-08-18 it leads the Artificial Analysis text-to-video (with audio) Elo table and is officially ~$0.10/s.
What it is
A preview multimodal video model (gemini-omni-flash-preview) that turns text, images, and short video into video with audio and then iterates via natural-language edits. Google markets it as "Nano Banana, but for video." From 2026-08-18-autoresearch-best-ai-video-generation-tools.
Why it matters to ai-video-generation
It is the current default general video tool on independent Elo and official price — not an avatar puppet, but the generate/iterate layer sitting next to this thread's heygen-avatar-v / runway-act-two / wan-animate lane. Google's product page also claims a VBench I2V human-pref tie with Grok-Imagine-Video and Kling (vendor eval, not AA).
Key facts
- Vendor: Google DeepMind / Gemini API.
- Status: public preview (pricing page, 2026-08-18 fetch).
- Official effective price: ~$0.10 per second of 720p (5,792 output tokens/s at $17.50 / 1M video tokens).
- AA T2V with audio (fetched 2026-08-18): Elo 1,239 ±7, 14,032 samples, $6.00/min — tied for first with minimax-h3.
- AA I2V with audio: Elo 1,187, third behind seedance-2-0 and minimax-h3.
- Output on the DeepMind page: generate + multi-turn edit; SynthID + C2PA on Gemini app / Flow / YouTube outputs.
Strengths (from our perspective)
- Highest independent T2V-with-audio Elo among models on the 2026-08-18 AA table, at a mid-pack price.
- Conversational editing is the product, not an afterthought.
- ~4× cheaper than veo-3-1 Standard at 720p/1080p on the same official price list.
Weaknesses (from our perspective)
- Closed / preview; not a pose-driven avatar model.
- Google's MovieGenBench / VBench numbers are vendor human evals.
- Clip-length and max resolution on the official product page are thinner than Wan3.0's 30 s claim.
Open questions
- How Omni Flash compares to wan-animate / runway-act-two on identity lock and hands — AA does not measure that.