Google Gemini 3.5 Flash/Omni agentic & creative push
Key Questions
What new agentic features were announced for Gemini 3.5 Flash at I/O 2026?
Gemini 3.5 Flash introduces agents for coding, Android integration, and Spark workflows. Google also launched Managed Agents in the Gemini API that enable single-API-call deployment in hosted sandboxes.
What capabilities does Gemini Omni Flash provide for video and simulation?
Gemini Omni Flash supports any-to-any video generation using Veo4 with Flow consistency and action replays. It also powers Genie simulations and achieves 93.3% on Deep Research tasks.
When is Gemini Spark available and to whom?
Gemini Spark is now available to all Google AI Ultra subscribers in the U.S. Hands-on reports indicate it delivers practical workflow value beyond initial hype.
What open-source contributions has DeepMind made for scientific agents?
DeepMind open-sourced scientific agent skills covering genomics, structural biology, and cheminformatics. The LEAP system demonstrates that scaffolding and verifier feedback can raise math reasoning from under 10% to 70% on Lean-IMO-Bench.
What image and branding tools were recently released by Google?
Nano Banana 2 image models are now generally available with video input support. Pomelli is a new branding tool that leverages these multimodal capabilities.
I/O 2026: Gemini 3.5 Flash agents/coding/Android/Spark; Omni Flash any-to-any video (Veo4, Flow consistency, action replays); Genie sims; Deep Research 93.3%. New: Nano Banana 2 image models GA with video input; Pomelli branding tool; SRE AI agents in production. Gemini 3.5 Flash crowned fastest model. Gemini Spark now available to all Google AI Ultra subscribers in the U.S. Hands-on shows real workflow value vs hype. New: Google Managed Agents in Gemini API—single API call, hosted sandbox for simplified agent deployment. Also: DeepMind open-sources scientific agent skills for genomics, structural biology, cheminformatics. Also: Google LEAP system shows scaffolding and verifier feedback can jump math reasoning from <10% to 70% on Lean-IMO-Bench, beating specialized systems.