Google AI Updates: October 7, 2026

1. Nano Banana 2.1 Ships in the Gemini API and Deprecates gemini-3.1-flash-image

Google. Google released gemini-nano-banana-2.1, an image generation and editing model built on Gemini 3.6 Flash, with improvements in prompt adherence, character consistency, and text rendering, plus extreme aspect ratios (1:4, 4:1, 1:8, 8:1) at 1K, 2K, and 4K. The model card lists a 1M token input context, up to 14 reference images, and availability in the Gemini app, AI Studio, the Gemini API, Search AI Mode, Flow, Stitch, and Google Ads; The Decoder reports per-image prices roughly halved to $0.0336 at 1K and $0.0756 at 4K. The previous gemini-3.1-flash-image model is now deprecated, so teams using it should plan a migration and re-test prompts that depend on text rendering or multi-character edits. Source

2. EmbeddingGemma 2 Adds Image, Video, and Audio to an Open 740M Embedding Model

Google. EmbeddingGemma 2 maps text, code, images, video, and audio into one embedding space, with a 270M text-only core and optional 170M vision and 300M audio encoders (740M total), 768-dimension vectors truncatable to 512, 256, or 128 via Matryoshka learning, and an 8K token context. Google reports 78.68 on MTEB Code, up nearly 10 points from the first version, with multilingual text quality unchanged and leading sub-1B results on the MAEB audio and MIEB image benchmarks. It is Apache 2.0 licensed, available on Hugging Face and Kaggle with support in Ollama, llama.cpp, MLX, and vLLM, and runs in about 191MB of RAM for text-only quantized weights on a Pixel 11 Pro, which makes on-device multimodal retrieval practical. Source

3. Google Research Maps Open Privacy and Security Problems for AI Agents

Google Research. A technical report with more than 50 academic and industry contributors applies Contextual Integrity theory to agents, judging not only which information flows are appropriate but whether an agent’s actions are appropriate in context. It recommends dynamic, context-aware sandboxing over static permissions, a contextual policy engine that supervises actions before execution, guardrails against multi-agent collusion, and standardized multi-agent benchmarks in simulated “Agent Gym” environments. For agent builders, it is a useful checklist of where prompt injection, probabilistic control flow, and delegated autonomy break traditional permission models. Source

4. Earth AI’s Population Dynamics Model Applied to Five Public Health Problems

Google Research. Google evaluated its Population Dynamics Foundation Model, which encodes aggregated search trends, mobility, built environment, and weather and air quality into monthly place embeddings, as a drop-in feature set for epidemiological models. Reported gains include a 36 percent relative increase in explained variance for MMR vaccination rates across 146 counties and an 18.1 percent improvement in Precision@5 for cholera emergence forecasting eight weeks ahead in the Democratic Republic of Congo. The embeddings are in preview on Google Maps Platform as Population Dynamics Insights and free for academic research by request, with an accompanying arXiv paper. Source