AI News from Google DeepMind Blog

Latest coverage from Google DeepMind Blog, summarized and scored for signal.

• AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome — Google DeepMind pre-computed molecular impact predictions for all 9 billion possible single-nucleotide mutations in the human genome — a 1-petabyte dataset 30× larger than AlphaFold — and released it free to academic researchers as AlphaGenome Atlas.
• Introducing WeatherNext 3, our most advanced and accurate global weather AI model — Google's WeatherNext 3 delivers hourly weather forecasts at 5km resolution — 5x sharper than WeatherNext 2 — using real-time satellite data, with precipitation accuracy up to 50% better at 1-day-plus horizons. Now live across Search, Maps, Gemini, and Google Cloud.
• Proactive cyber defense for governments and enterprises — Google DeepMind launches Fairwind Program, giving governments and critical infrastructure operators access to Gemini 3.8 Flash Cyber and the CodeMender harness to autonomously detect and generate deployment-ready vulnerability patches within their own secure cloud environments.
• Introducing Gemini 3.8 Flash and 3.8 Flash Cyber — Google DeepMind releases Gemini 3.8 Flash at the same price as 3.7 Flash ($0.75/M input, $3.75/M output), with substantial gains in coding and agentic tasks — including 54.9% on HLE-Verified and top scores on the DeepSWE v1.1 long-horizon software engineering benchmark, often surpassing larger frontier models.
• Introducing agentic video understanding with Gemini — Google DeepMind launched agentic video understanding for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, replacing fixed-rate frame ingestion with dynamic scanning across video, audio, and transcripts — cutting token usage by up to 88%, costs by up to 66%, and boosting accuracy by up to 7% on benchmarks.
• Gemini Omni 1.1 Flash lets you build with more control — Google DeepMind released Gemini Omni 1.1 Flash to the Gemini API, adding production-ready video generation controls for developers: scene extension with 10x more prior context (10s vs. 1s), first/last frame interpolation for camera control, 4K upscaling, and 360p preview mode to cut iteration costs.
• Piloting the world's first double-blind AI evaluations — Google DeepMind and partners including Singapore's AI Safety Institute ran the first cryptographically enforced double-blind evaluation of a frontier AI model, using Confidential Computing to prevent both benchmark contamination and IP exposure — a direct technical response to growing distrust of self-reported AI benchmarks.
• Intelligent transcription with Gemini 3.5 Transcribe — Google DeepMind launched Gemini 3.5 Transcribe, achieving 2.6% word error rate for pre-recorded audio and 4.0% for real-time streaming — now available via the Gemini API with speaker attribution, word-level timestamps, and 85-language support for developers building voice agents and call analytics pipelines.
• Introducing Gemini 3.7 Flash — Google DeepMind officially launched Gemini 3.7 Flash — three weeks after 3.6 Flash — with FrontierCode jumping from 34.4% to 43.6%, DeepSWE from 49.0% to 65.3%, and AutomationBench from 17.0% to 30.4%, at $0.75/1M input tokens locked through year-end.
• Putting sign language AI into users’ hands — Google DeepMind's SL2T model ships in Pixel 11, enabling Deaf users to sign instead of type anywhere on their phone — the first consumer deployment of sign language AI at this scale, starting with ASL and expanding to additional languages.
• WeatherNext: AI model achieves breakthrough in forecasting cyclones — DeepMind open-sourced WeatherNext, a Nature-published AI model that unifies hurricane track and intensity prediction, delivering one extra day of average lead time — equivalent to a decade of meteorological progress — and predicting 1,000 storm scenarios per cyclone.
• Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration — Google DeepMind released Gemini Robotics ER 2, an embodied reasoning model now available via the Gemini API that acts as a high-level robot brain — enabling real-time video understanding, multi-step task planning, and multi-robot collaboration.
• We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control — Google DeepMind launches Lyria 3.5 in Google Flow Music, improving across all four core dimensions: richer melodic structures, higher-quality lyrics with structural awareness, more expressive and emotionally nuanced vocals with better pronunciation, and direct tempo and duration controls.
• Gemini Robotics 2 brings whole body intelligence to robots — Google DeepMind has released Gemini Robotics 2, a suite of three AI models enabling whole-body humanoid control (from feet to fingertips), advanced dexterous manipulation, multi-robot collaboration, and fast adaptation to new robot bodies in hours — not weeks.
• Accelerating the frontiers of scientific discovery: Google’s $40M commitment to the Genesis Mission — Google is committing $40M in AI tokens and cloud credits to the DOE's Genesis Mission, granting all 17 Department of Energy National Laboratories access to AlphaFold 3, AlphaEvolve, AlphaGenome, WeatherNext, and Gemini for Government seats — the largest Google AI-for-science commitment to a federal research program to date.
• Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — Google released three new Gemini models — 3.6 Flash (17% fewer output tokens than 3.5 Flash, priced at $7.50/1M output), ultra-fast 3.5 Flash-Lite at 350 tokens/second, and 3.5 Flash Cyber for cybersecurity — while confirming Gemini 4 pre-training has begun.
• Introducing Gemini 3.5 Flash Cyber — Google introduced Gemini 3.5 Flash Cyber, a lightweight cybersecurity model fine-tuned on top of 3.5 Flash, designed for repeated invocation to scan large codebases — offering competitive vulnerability detection against larger models while available exclusively to governments and trusted partners via CodeMender.
• Our approach to bioresilience — Google DeepMind and Isomorphic Labs have published a joint bioresilience framework — covering SynthID watermarking adapted for AI-generated DNA sequence screening, AlphaEvolve-powered pathogen surveillance, and 15+ government biosecurity partnerships — marking the first comprehensive public biosecurity program from a frontier AI lab.
• Empowering India’s next generation of innovators with ATL Saathi — Google DeepMind and India's Atal Innovation Mission launched ATL Saathi, a Gemini-powered assistant for Atal Tinkering Lab educators serving 11 million students — marking Gemini's first major government-backed education deployment in India.
• Google DeepMind and A24 announce first-of-its-kind research partnership — Google DeepMind and A24 announced a long-term R&D partnership to co-develop AI tools for filmmakers, with Google taking an investment stake in A24—the first tie-up between a frontier AI lab and a filmmaker-led studio, though no specific technology, timeline, or model has been announced yet.
• Start building with Nano Banana 2 Lite and Gemini Omni Flash — Google released Nano Banana 2 Lite (gemini-3.1-flash-lite-image), a 4-second, $0.034/1K-image model for high-throughput pipelines, alongside Gemini Omni Flash — now accessible to developers via the Gemini API for video generation and multi-turn conversational editing.
• Introducing computer use in Gemini 3.5 Flash — Google has integrated computer use directly into Gemini 3.5 Flash — previously a standalone Gemini 2.5-only capability — enabling agents to see, reason, and act across browser, mobile, and desktop via the standard Flash API, with enterprise safeguards including automatic prompt injection detection.
• Unlocking UK house-building with AI-accelerated planning — Google DeepMind is partnering with the UK government to deploy a Gemini-powered planning tool targeting a 50% cut in householder application decision times, with national rollout to all UK councils planned from 2027.
• Securing the future of AI agents — Google DeepMind published its AI Control Roadmap — a defense-in-depth security framework for managing potentially misaligned AI agents inside Google, built on the MITRE ATT&CK threat model and treating internal agents as potential insider threats even after alignment training.
• DiffusionGemma: 4x faster text generation — Google DeepMind released DiffusionGemma under Apache 2.0 — a 26B MoE model that generates 256 tokens in parallel using text diffusion, hitting 1,000+ tokens/sec on H100 and 700+ on RTX 5090, up to 4x faster than autoregressive models, with immediate Hugging Face, vLLM, and Unsloth support.