Gemini Omni 1.2 Flash hits GA: 15-second video, native audio

Google's video model is now in the paid Gemini API: 720p with synced audio, ~$1 a clip — and a Dec 3 auto-migration that quietly reroutes Omni 1.1 calls.

Nowline NOV 5 9:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Text-to-video with sound, straight from the API

    GA covers 3–15 second clips (up from a 10-second cap) at 720p with synchronized native audio and multi-turn conversational editing that keeps the prior clip server-side, so you don't re-upload it. Prototype free in Google AI Studio; production runs on the paid Gemini API. You can extend a clip 3–10s at a time up to 30s total, and upscale selected drafts to 1080p or 4K.

  • What a clip actually costs

    Video output meters at $12 per million tokens — 5,792 tokens per second of 720p — so a 10-second clip runs about $0.70 and a 15-second clip about $1.04; the 720p rate was cut 30% to roughly $0.07 a second. Input (text, image, or video) is $1.50/M and there's no free tier, so budget before you wire this into an agent loop.

  • The Dec 3 gotcha: your 1.1 calls silently reroute

    Omni 1.1 Flash stays live for a four-week transition, then on December 3 every API request to 1.1 auto-routes to 1.2 with no code change — and bills at 1.2 rates. If you pin model versions in production, test against 1.2 now so new output behavior and new costs don't surprise you on migration day.

  • Reference images in, English-only out

    Up to six reference images per request (three per character, for two characters) give you identity-consistent characters across shots; DeepMind claims a 63% vs 37% win rate over 1.1 on multi-reference consistency. The catches: the API is English-only for now, audio references are allowlist-gated, and fast sports action and complex physics still break.