The latest AI news and updates of Google — the company behind Gemini, DeepMind research, Gemma open models, and AI across Search, Cloud, and Workspace. Covering Google's latest product updates, launches, and research from the past 90 days.
Gemini
Google Gemma
Google AI Studio
Google13h agoGoogle Research partnered with Ignite Energy Access to build a repeatable blueprint for scaling off-grid solar across sub-Saharan Africa. The project uses Gemini for Science and the Open Buildings dataset to map remote homes and optimize logistics. This collaboration enabled an estimated 160,000 tCO2e reduction in H1 2026, with plans to model micro-grids for clean cooking.
Google Gemma23h agoGoogle Gemma reports a 2.85x inference speedup for the Gemma 4 E2B model on A100 GPUs via vLLM. The optimization, implemented using Gemini 4 Argon, increases throughput from approximately 37,900 to over 108,000 tokens per second. These changes, delivered across eight upstream pull requests, maintain 100% token parity and zero quality loss.
GoogleOct 26Google Research is demonstrating its SL2T model at EMNLP 2026, which powers ASL-to-English sign-to-text translation in Gboard and Live Transcribe on Pixel 11. The model translates sign language directly into text by tracking body pose landmarks, bypassing intermediate gloss annotations. This feature is available now on Pixel 11, with support for additional languages and devices planned.
GoogleOct 26Google Research released DeepSEA, a benchmark evaluating LLMs across 12 Southeast Asian languages. The project covers five tasks, including mathematical and scientific reasoning, using human-translated datasets. Findings indicate that while frontier models maintain robustness, smaller models struggle with low-resource languages. Shubham Mittal is demoing the benchmark today at 10:30am at the EMNLP 2026 Google booth (#D5).
GoogleOct 24Google Research is showcasing how open foundation models like MedGemma support TB screening at the point of care. Partner Nexus AI has deployed certified diagnostic tools across six countries, reducing turnaround times to under 45 seconds. The initiative includes acoustic triage via HeAR and supports national programs like Indonesia’s goal of screening 50 million citizens annually.
Google GemmaOct 23Google Gemma reports that the community fine-tuned EmbeddingGemma 2 on 3,000 hours of Vietnamese audio in just 3 hours of training. The resulting 356M-parameter model achieves the lowest average word error rate of 6.87% across eight Vietnamese speech recognition benchmarks, outperforming models like Gemini-Flash-3.8 and Whisper without requiring a heavy autoregressive decoder.
GoogleOct 23Google Research published a Nature paper detailing the Personal Health Agent (PHA), the multi-agent architecture behind Google Health Coach. The system orchestrates specialized agents to analyze wearable and biomarker data for personalized health insights. Google Research has now open-sourced the framework, providing the community with access to the underlying architecture.
GeminiOct 22Google Gemini now references Google Maps searches and saved places alongside Google Calendar to provide tailored daily recommendations. This integration functions through the Personal Intelligence setting, where Gemini synthesizes location history and schedule data to assist with planning. The feature requires explicit permission to connect these apps to the Gemini experience.
GoogleOct 22Google Research completed its 32-bit Coral NPU architecture, adding a Matrix Execution Unit that delivers up to 1 TOPS at 10 mW. The release includes a v1.0 IREE/MLIR compiler and new Synaptics Astra SR200 chips. These tools allow model optimization for deployment across diverse, battery-powered edge devices like wearables and industrial sensors.
Google DeepMindOct 22Google DeepMind, with Google Research, WRI, and IIASA, released Forest Typology (ForTy), a 10-meter resolution map categorizing global land into six forest and land-use types. The release includes an annual 2017–2025 time series tracking land-use changes. Both datasets are available on Google Earth Engine to help organizations build deforestation-free supply chains and protect natural forests.
GoogleOct 21Google Research completed its 32-bit open-source Coral NPU architecture, adding a Matrix Execution Unit for up to 1 TOPS at 10 mW. The team also released the v1.0 IREE/MLIR compiler for JAX, PyTorch, and TensorFlow models. Additionally, Synaptics launched the Astra SR200 series and new Coralboard devkits, supporting optimized model deployment across diverse edge hardware.
GoogleOct 21Google Research published a prospective study in JAMA Network Open demonstrating that its blind sweep ultrasound AI enables non-sonographer operators to estimate gestational age with a mean absolute error of 4.18 days. Tested across Chicago and Nairobi using low-cost handheld probes, the model achieved non-inferior accuracy to expert sonographers, who recorded a 4.48-day mean absolute error.
Google GemmaOct 21Google released Google AI Edge Foresight, an experimental macOS app that provides offline meeting assistance. Powered by the 740M-parameter EmbeddingGemma 2 model, the app auto-enhances shorthand notes and performs semantic search across local documents, images, and transcripts. A live assistant detects questions during meetings and retrieves answers directly from local files without requiring an internet connection.
GoogleOct 20Earth Fire Alliance, in partnership with Google Research and Muon Space, launched the first FireSat satellites to detect wildfires from orbit. The system uses AI trained on controlled burns to identify fires as small as 5x5 meters. By 2030, the 50-satellite constellation will capture a complete picture of the Earth’s surface every 20 minutes.
Gemini NotebookOct 20Google Gemini Notebook introduces a co-creation flow for Slide Decks and Video Overviews. A new review stage provides an editable outline or producer notes for refinement before final generation. This process incorporates specific feedback on topics, formats, and visual styles. The feature is available to Pro and Ultra subscribers, with a global rollout occurring over the next week.
Google DeepMindOct 17Google DeepMind’s Gemini 4 Argon agents defeated Minecraft’s Ender Dragon from scratch in 27 hours. The four-agent team autonomously wrote 2,783 programs, coordinated tasks through chat, shared notes, and debugged errors without human intervention. The agents operated in a sandbox using only code and text to navigate, mine, and survive until the final objective was reached.
GoogleOct 16Google Research is grounding AI agents in physical lab robotics to run closed-loop molecular design cycles. The team introduced the VERID framework for lab state tracking and demonstrated autonomous discovery of single-digit nanomolar protein binders for cancer targets. These agents iteratively refine experiments, achieving significant hit-rate improvements in aptamer and de novo protein design campaigns.
GoogleOct 15Google AI demonstrated Gemini 4 Argon acting as a virtual builder within the Google Antigravity platform. The model autonomously constructed a fully-explorable, multi-tiered pagoda sanctuary in minutes. Argon uses long-horizon autonomy and spatial reasoning to manage 3D geometry, lighting, and physics, allowing the model to execute complex architectural builds from a single prompt.
GoogleOct 14Google Research launched Deep Review, an agentic system for rigorous document verification. Originally used for 16,000+ academic submissions at conferences like NeurIPS and ICML, the tool now supports Google Docs and presentations. It uses a four-stage agentic pipeline to check logical consistency and ground claims. The system provides two complimentary reviews weekly, with API access available for high-volume workflows.
GoogleOct 14Google Research published a technical report on its Geospatial Reasoning agent, which combines domain-specific foundation models, spatial tools, and upfront planning. The system performs complex spatial analysis and planetary prediction that standalone LLMs cannot achieve. Google is now expanding experimental access to the agent through Ask Google Earth and an Agent-to-Agent API for custom developer workflows.
Google DeepMindOct 14Google DeepMind and GovTech Singapore are building an on-device scam detection model using the open Gemma model. Fine-tuned on local Singaporean scam tactics, the system analyzes conversation patterns in real time to flag manipulative intent. This approach keeps all data on the device, ensuring personal privacy while protecting against social engineering schemes that account for over half of local crimes.
Google DeepMindOct 13Google DeepMind’s Gemini 4 Argon taught Apptronik’s Apollo humanoid to play cornhole by watching a video. The model mapped the physics, generated its own simulation, and guided the robot to sink seven consecutive throws. This approach compresses weeks of manual modeling into just a few hours. Gemini 4 Argon is now available in Google AI Studio and the Gemini API.
GoogleOct 13Google launched Gemini 4 Argon, a frontier model designed for complex tasks in coding, finance, legal, cybersecurity, and science. The model features a 1M output token limit and is available globally via the Gemini API, Antigravity, and the Gemini app. Google also opened applications for scientists to access Antigravity and for cyber defenders to join the guardrail-free Fairwind Program.
Google GemmaOct 9Google AI Pro and Ultra plans now include premium Colab compute, providing subscribers with A100 80GB or H100 GPUs. This integration supports browser-based execution and fine-tuning of Gemma 4 models, including full fine-tuning for the E4B variant and QLoRA tuning for the 31B model in bf16 precision.
GoogleOct 9Google Research and Beth Israel Deaconess Medical Center published a study in The Lancet evaluating the AMIE medical AI in a real-world clinic. Across 100 patient interactions, AMIE recorded zero safety stops and matched doctor diagnoses in 90% of cases. The study marks Google's first publication in the journal's main section, with clinicians reporting AI summaries helped prepare for visits.
GoogleOct 8Google Research conducted a three-month randomized controlled trial with 133 patent attorneys using an AI writing assistant. While AI access improved drafting productivity for all, unassisted judgment gains varied by experience. Senior lawyers outperformed controls by 0.45 standard deviations, while junior lawyers showed no average skill improvement, with their performance scores bifurcating into more high and low results.
Google DeepMindOct 8Google DeepMind released the SynthID Detector to the public, providing a portal to verify digital watermarks in images, video, and audio. The tool identifies AI-generated content from Google, OpenAI, NVIDIA, and Kakao, with Apple support arriving soon. It detects watermarks even after modifications like filters or compression, confirming if media originated from these partner models.
Google LabsOct 7Google Labs launched Playground, an experimental gaming platform that generates custom games from text prompts. The platform requires no coding experience and offers genres like trivia, tower defense, and racing. Available to US users 18 and older, the chat-based interface enables the creation, play, and sharing of custom games.
AntigravityOct 7Google Antigravity demonstrated an agentic workflow that takes an Android app from a natural language prompt to a physical device. The agent uses the Stitch MCP server to pull designs and the Android CLI plugin to build native Jetpack Compose components, verify them in an emulator, and execute the final build on a connected phone.
GeminiOct 6Google launched Guided Vision in Gemini Live, providing real-time visual assistance for blind and low-vision users. The feature uses camera input to generate dynamic audio descriptions and spoken reframing cues, such as panning or tilting instructions. It is available on Android 9 and above in regions where Gemini Live is supported, with integration via accessibility shortcuts and TalkBack.
Google GemmaOct 6Google released EmbeddingGemma 2, a 740M-parameter open model that maps text, code, images, video, and audio into a unified embedding space. Built on the Gemma 4 architecture, it features an 8K context window and delivers a 9.92-point improvement on code performance over the original model. The Apache 2.0 model is optimized for on-device RAG and semantic search.
Google AI StudioOct 6Google AI Studio launched Nano Banana 2.1, an upgraded image generation and editing model. This version delivers improvements in visual design, mask-based editing, subject consistency, and image naturalness. The model is available for use today within the Google AI Studio environment.
GoogleOct 5Living Models paired Google’s Gemma 4 with its specialized BOTANIC-1 model to accelerate plant DNA analysis. In a test on a melon yield gene, the pipeline identified the causal mutation as the top candidate out of 2,494 possibilities in under four minutes. This modular approach replaces years of physical crop breeding with rapid computational screening for climate-resilient agriculture.
GoogleOct 2Google launched a prototype satellite on SpaceX’s Transporter-18 mission to test how its Tensor Processing Units perform in space. Developed with Planet, the mission gathers data on hardware survival against radiation, vibration, and vacuum cooling challenges. This test informs future plans to link satellite constellations for scalable machine learning, leveraging near-constant solar power in low Earth orbit.
GoogleOct 2Google Research introduced a next-generation Federated Learning system that uses Trusted Execution Environments to provide verifiable differential privacy. By shifting computation server-side, the system reduces training times and improves model accuracy. Gboard has already deployed this infrastructure for English and Japanese next-word prediction, with all server-side workloads and privacy-preserving logic published to a public transparency log.
Google GemmaOct 1Google reports that its Gemma 4 31B model achieves GRE learning gains statistically equivalent to expert human tutors. Research involving 2,383 participants found the model matched human performance with a p-value of .015. The AI tutoring cost $0.0052 per percentage point gained, a 918-fold reduction compared to the $4.81 cost of human instruction.
Google DeepMindOct 1Google DeepMind introduced SynthID Bio, a watermarking technology that embeds imperceptible signatures into AI-generated protein sequences and structures. The method preserves biological function, allowing DNA synthesis providers to verify the provenance of synthetic proteins. Google DeepMind released the code, in vitro data, and model weights on an open basis for research use to strengthen biosecurity and scientific integrity.
GoogleOct 1The CDC identified Google’s science AI model as the top-performing entry among 39 eligible models for forecasting U.S. flu-related hospital admissions during the 2025-26 season. Developed using Google’s Empirical Research Assistance, the model matched observed hospital admissions throughout the season. The underlying technology is currently available to trusted testers through Google’s experimental science tools.
GoogleSep 30Google launched Gemini 4 Argon, a frontier model designed for deep reasoning in software engineering, enterprise knowledge work, and cybersecurity defense. The model features an expanded 1M output token limit to support long-horizon tasks. Argon is currently rolling out to trusted cyber defenders through the Fairwind Program, with broader availability planned for developers and enterprises soon.
GeminiSep 30Google is rolling out Skills in Gemini globally today, providing reusable, stackable instruction sets that replace Gems. Skills save custom instructions, support reference files like PDFs or images, and trigger via a forward slash. Existing Gems migrate automatically, with support for Gems ending in November for personal accounts and through 2027 for Workspace business and education customers.
GoogleSep 29Google Research introduced Diffusion Controller, a lightweight steering-damper network that improves prompt alignment in image generation without breaking baseline stability. The framework attaches to frozen base models, including closed-source systems, to dynamically adjust generation trajectories. In testing, the white-box version achieved a 90% win rate over baseline models, outperforming standard parameter-efficient fine-tuning methods like LoRA.
GoogleSep 28Google expanded its experimental CC AI agent into a group assistant for families. CC now holds its own verified Google account, providing a unified view of shared emails, documents, and images. The agent manages household logistics by updating shared calendars, pre-filling registration forms, checking live drive times, and reconciling grocery lists. Access is available via a waitlist.
AntigravitySep 26Google Antigravity 2.0 now includes a dedicated planning mode, bringing the IDE to parity with the Antigravity CLI. Invoking the /plan command triggers an agentic research phase that generates a reviewable implementation plan. The agent pauses execution for user approval, ensuring complex tasks are scoped and verified before any code modifications occur.
GoogleSep 25Google Research introduced a unified multi-agent framework that autonomously generates temporally consistent, long-form video narratives. The system uses hierarchical orchestration to mitigate visual drift and pipeline error propagation across multi-shot sequences. It includes four specialized components—Co-Director, CANVAS, A²RD, and VQQA—to manage creative intent, visual memory, temporal dynamics, and closed-loop refinement, achieving an 81.4 quality score on GenAD-Bench.
AntigravitySep 24Google Antigravity SDK now supports offline local execution for Gemma 4 models using Google AI Edge's LiteRT. This update allows agentic workflows to run entirely on local GPUs, eliminating API costs and keeping data on-device. The SDK also adds support for OpenAI-compatible backends like Ollama and vLLM, providing flexibility for local inference configurations.
Gemini NotebookSep 23Google integrates Gemini Notebooks as a context source within Google Docs through Workspace Intelligence. This update combines deep research from notebooks with existing emails, chats, and files to personalize drafts. The integration allows direct referencing of notebook content to inform writing processes within the document interface.
GeminiSep 23Google Gemini integrates 13 new third-party services, including Adobe, Airtable, Linear, and Peloton, via the Model Context Protocol. These connections support direct management of projects, design assets, and fitness planning within the Gemini interface. The update rolls out today, with integrations accessible through Gemini settings or by mentioning the apps directly in chat.
GoogleSep 23Google launched Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, two audio models supporting 100+ languages and 2,000+ voices. Flash TTS enables bespoke voice design and line-by-line performance direction, while Flash-Lite TTS optimizes for real-time voice agents. Both models are available today in the Gemini API and Google AI Studio, with specific integrations in Gemini Notebook and Google Vids.
Google GemmaSep 23Google now supports single-command deployment for DiffusionGemma-Jev on Google Cloud Run. This setup provides a Jev API-compatible endpoint with single-step latency of 35-60 ms and batch throughput of 100-123 requests per second. The service costs approximately $3 per hour and scales to zero when idle, removing the need for local GPU hardware during experimentation.
GoogleSep 21Google AI Developers demonstrated a coding tutor powered by Gemini 3.8 Live Extended Thinking. The agent uses real-time screen awareness and function calling to reference the p5.js library, narrating its logic while debugging code. This model supports multi-step reasoning and background tool execution, allowing the AI to manage tasks without interrupting the conversational flow.
Gemini NotebookSep 21Gemini Notebook has fully rolled out Live Chat to all Ultra users on mobile. The feature supports real-time voice conversations across nearly 100 languages, providing hands-free, step-by-step guidance on source material. Powered by the latest audio models, the interface also supports mid-conversation language switching.
Gemini NotebookSep 21Google Gemini Notebook has made Interactive Learning Overviews available to all users. Located under Reports, the feature merges source summaries with studio artifacts like infographics, quizzes, and flashcards into a single interactive hub. This update completes the rollout of the tool, which organizes study materials into a centralized, personalized learning experience.
GoogleSep 17Google Research introduced a generative UI framework that creates custom, guided educational simulations for STEM topics. The system uses pedagogical guardrails to ensure instructional alignment and accuracy. A public library of over 30 STEM interactives is now available, and schools using Google Workspace for Education can join a pilot program to request and validate new simulations.
Google AI StudioSep 17Google AI Studio and Speakeasy open-sourced the Speakeasy OpenAPI client suite under the AGPLv3 license. The suite includes multi-language SDK generators for seven languages, an agent-native CLI generator, and a documentation MCP server generator. This release follows Google’s migration of its GenAI SDKs after its previous proprietary generator provider shut down in May 2026.
Google LabsSep 17Google Labs launched a new CC AI agent for families, supporting up to five members. The agent synchronizes shared Google Calendar and Tasks, coordinates logistics in Google Chat, and delegates paperwork. It also sends daily morning briefings to the group. Access is available via waitlist or upgrade for US residents aged 18 and older.
Google GemmaSep 16Google updated the AI Edge Gallery app on macOS to include full support for the Gemma 4 12B model. The update introduces multimodal vision and audio input processing, configurable vision token budgets, and MTP for high-speed inference. These capabilities are optimized to run locally on 16GB Mac hardware, enabling private, offline generative AI workflows.
GeminiSep 16Google Gemini now enables users to code custom applications within Canvas to design 3D objects. This update includes 3D visualization and the ability to export designs as .STL files for 3D printing. Canvas is available to all Gemini users, with Gemini 3 and a 1 million token context window provided for Pro and Ultra subscribers.
AntigravitySep 16Google Antigravity updated its Terminal Sandbox to run shell commands automatically within a secure, project-isolated environment. This system restricts network access by default, reducing the need for manual command approvals. The update is available today on macOS and Linux, enabling agents to perform more tasks out of the box while maintaining workstation security.
GoogleSep 15Google Research introduced Retrieve-for-Train, a framework that accelerates complex AI search by replacing autoregressive inference with a lightweight diffusion model. By using offline reinforcement learning to train a 53.9M-parameter model, the system generates diverse, grounded sub-queries in a single pass. This approach delivers a 12 to 20 times speedup over traditional methods, bypassing the need for extensive reasoning tokens.
GoogleSep 15Google released a white paper detailing its AI models for predicting natural disasters, including riverine floods, urban flash floods, cyclones, and wildfires. The company uses technologies like Groundsource and WeatherNext to provide early warnings to partners in over 100 countries. These models extend prediction horizons, helping communities prepare for extreme weather events and mitigate disaster impacts.
Frequently asked questions