Google releases Gemini Omni Flash video generation model with conversational editing, withholds speech synthesis
Google DeepMind released Gemini Omni Flash, the first model in its new Omni family that generates and edits video from image, audio, video, and text inputs. The model is rolling out to Gemini app subscribers and YouTube Shorts with a 10-second clip limit, while speech-editing capabilities remain withheld pending safety testing.
Gemini Omni Flash — Quick Specs
Google releases Gemini Omni Flash video generation model with conversational editing, withholds speech synthesis
Google DeepMind released Gemini Omni Flash on Tuesday at I/O 2026, a multimodal video generation model that accepts any combination of image, audio, video, and text inputs. The model began rolling out the same day to Gemini app users with AI Plus, Pro, and Ultra subscriptions, YouTube Shorts, and the YouTube Create app.
Model capabilities and limits
Gemini Omni Flash generates video clips capped at 10 seconds at launch. According to Koray Kavukcuoglu, CTO of Google DeepMind, the model supports conversational editing where each instruction builds on previous ones while preserving character identity and scene continuity across multiple turns.
Google claims the model has improved physics simulation including gravity, kinetic energy, and fluid dynamics. The company demonstrated prompts ranging from claymation protein-folding animations to chain-reaction physics tracks, though benchmark scores comparing Omni to Veo 3 or competing models like ByteDance's Seedance have not been disclosed.
The 10-second limit is shorter than OpenAI's Sora, which generates clips up to 60 seconds. Google has not disclosed per-clip costs, compute footprint per generation, or pricing for tiers beyond Flash.
Speech editing withheld, SynthID mandatory
Google is explicitly withholding general-purpose audio and speech editing capabilities. "We are still working to test this and better understand how we can bring this capability to users responsibly," Kavukcuoglu wrote, in what appears to be a deliberate step back from deepfake territory.
The model includes digital avatar generation requiring users to record their voice and likeness by speaking a series of numbers aloud during onboarding.
All videos generated with Omni carry Google's SynthID imperceptible digital watermark by default. Users can verify whether a clip was generated by Omni through the Gemini app, Gemini in Chrome, and Google Search. The SynthID implementation follows the C2PA open standard that OpenAI adopted earlier this year.
Availability and rollout
API access for developers and enterprise customers will launch in the coming weeks. Google has not disclosed the underlying model architecture relative to Veo 3, benchmark evaluation methodology, or timeline for enabling speech editing across the Omni family.
The model launched alongside Gemini 3.5 and other announcements at I/O 2026 that Sundar Pichai framed as the "agentic Gemini era" in his keynote.
What this means
Google's decision to withhold speech editing while releasing video generation marks a conservative deployment strategy compared to frontier competitors. The 10-second clip limit and lack of disclosed pricing create uncertainty about whether Omni represents a technical advance or primarily a product integration of existing capabilities. The mandatory SynthID watermarking positions Google as prioritizing provenance over feature completeness, though the effectiveness of imperceptible watermarks against adversarial removal remains an open question. The API rollout in coming weeks will reveal whether the cost structure and extended clip lengths under paid tiers can compete with Sora's current specifications.
Related Articles
Mistral AI Releases Shieldstral-1.0-3B, a 3B-Parameter Policy-Adaptive Safety Classifier
Mistral AI has released Shieldstral-1.0-3B, a compact open-weight safety classifier that evaluates text and images against natural-language policies specified at inference time. The 3B model runs on a single GPU and reports F1 scores competitive with or exceeding larger moderation models like LlamaGuard-4-12B and GPT-OSS-Safeguard-20B on multiple benchmarks.
Mistral Releases Shieldstral, a 3B Open-Weights Safety Classifier That Matches Models 7x Its Size
Mistral has released Shieldstral, a 3B open-weights safety classifier that reframes content moderation as a policy-adaptive question-answering task. The model claims to match or outperform guard models up to 7x its size on text safety and multimodal benchmarks, and runs on a single 16GB GPU.
OpenAI Halts Parts of Astra Model Development After It Hit 'Critical' Cybersecurity Threshold
OpenAI disclosed that its in-development Astra model showed cyberattack capabilities strong enough that it cannot rule out a 'Critical' risk classification. The company has paused related internal activity and added security controls under its Preparedness Framework.
Mistral's 3B-Parameter Shieldstral Matches 20B Safety Model on Text Benchmarks
Mistral's new Shieldstral, a 3-billion-parameter open-weight safety classifier, posts an 84.9% F1 score on text benchmarks—tying OpenAI's GPT-OSS-Safeguard-20B, a model roughly seven times larger. The model lets operators define safety rules at runtime using plain-language yes/no questions instead of fixed taxonomies.
Comments
Loading...