Google

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ...Google search ...Google News ...Bing News


Google is one of the broadest artificial intelligence organizations in the world, and it is best understood as a family of companies rather than a single product team. Its parent company, Alphabet, combines Google (Search, Android, Chrome, YouTube, Workspace, Pixel, Google Cloud), the merged research lab Google DeepMind, Google Research, and a set of "Other Bets" and affiliates such as Waymo, Wing, Isomorphic Labs, Calico, and X, the moonshot factory. Google has been developing AI for more than two decades: products such as Google Lens and Google Translate were built with optical character recognition and machine learning, and AI has long powered core services such as the search engine, voice search, and Google Photos. Today nearly every Google surface runs on the Gemini model family, while specialized systems handle science, medicine, weather, robotics, and driving.

What distinguishes Google from most competitors is that it owns the entire AI stack. It designs its own chips (Tensor Processing Units, or TPUs), runs some of the world's largest data centers, trains frontier large language models and open models, and distributes them through products used by billions of people. It also operates one of the most productive scientific AI programs anywhere: AlphaFold, the protein structure system from Google DeepMind, earned Demis Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry. This combination of hardware, models, distribution, and science is unusual; model-focused labs such as OpenAI and Anthropic rely on partners for chips and consumer reach, while device makers such as Apple rely on partners for frontier models (Apple now licenses Gemini technology for Siri).

The scale of this effort grew sharply in 2025 and 2026. Google reported that the Gemini app passed 900 million monthly users by Google I/O in May 2026, and Sundar Pichai cited more than 950 million when he announced a leadership reorganization in August 2026. AI Overviews in Search reached roughly 2.5 billion monthly users, AI Mode passed one billion users in its first year, and Pichai told the I/O audience that capital spending in 2026 would be roughly $180 billion to $190 billion, more than six times the 2022 level. Those figures explain why Google describes 2026 as "the agentic Gemini era": the company is moving from chat answers toward AI agents that plan, act, and keep working in the background.

This page catalogs Google's AI offerings company by company and use case by use case: frontier and open models, consumer apps, Google Labs experiments, developer tools, Google Cloud, science, protein design and drug discovery, healthcare, climate and carbon tools, autonomous systems, education, and safety. A note on names helps here, because "Carbon" now means three different things at Google. Gemini 4 Carbon is the frontier model announced on October 28, 2026; carbon in the climate sense covers tools such as Project Contrails, the Environmental Insights Explorer, and Google's own emissions reporting; and Carbon is also the name of an experimental open-source programming language that Google engineers proposed as a successor to C++. Each is covered in its own section below.


Google's AI advantage comes less from any single model than from owning every layer at once: custom chips, frontier and open models, consumer distribution at the scale of billions, and a scientific research engine that turns the same techniques into Nobel-level discoveries, medical tools, and climate forecasts.


Alphabet's AI Organization

Alphabet's AI work is spread across several organizations with different missions, and knowing which unit builds what makes the product landscape easier to follow. Google DeepMind, formed in April 2023 by merging DeepMind (acquired in 2014 for roughly $500 million) with the Google Brain team, builds the Gemini, Gemma, Imagen, Veo, Lyria, and Genie models as well as the "Alpha" family of scientific systems. Google Research continues longer-horizon work in health, climate, quantum computing, and the foundations of machine learning, and it ships many of the climate and medical tools described later. Google Cloud packages these models for businesses, governments, and developers, and the product areas (Search, Android, Chrome, YouTube, Workspace, Devices and Services) put them in front of consumers.

Leadership changed significantly in 2026. On August 5, 2026, Alphabet announced that Demis Hassabis would step back from day-to-day leadership of Google DeepMind to become its chair and Alphabet's chief scientist, a role focused on long-horizon science, AGI, and safety, while continuing to lead Isomorphic Labs. Koray Kavukcuoglu, previously Google DeepMind's chief technology officer and Alphabet's chief AI architect, became senior vice president running Google DeepMind and reports directly to Sundar Pichai, and longtime engineer Jeff Dean left the company. The reorganization followed other high-profile departures, including AlphaFold co-creator John Jumper, who announced in June 2026 that he was leaving for Anthropic.

Organization Role in Alphabet Representative AI offerings
Google DeepMind Core AI lab (models and science) Gemini 4 Argon and Carbon, Gemini 3.x Flash, Gemma 4, Nano Banana, Veo, Gemini Omni, Lyria, Genie 4, Gemini Robotics 2, AlphaFold, AlphaGenome, AlphaEvolve, WeatherNext, SynthID
Google Research Applied and foundational research AMIE, MedGemma, Health AI Developer Foundations, Flood Hub, Project Contrails, Green Light, Co-Scientist, Earth AI, Quantum AI
Google (product areas) Consumer and developer products Gemini app and Gemini Spark, Search AI Mode, Chrome, Android and Pixel, Gemini for Home, Workspace, Gemini Notebook, Google Labs apps, Google AI Studio, Google Antigravity
Google Cloud Enterprise and public sector Gemini Enterprise, Gemini Enterprise Agent Platform (formerly Vertex AI), TPUs, AI Hypercomputer, Wiz, Mandiant, AI Threat Defense
Isomorphic Labs Drug discovery spinout of DeepMind Isomorphic Labs Drug Design Engine (IsoDDE), internal drug pipeline
Waymo Autonomous driving Waymo Driver robotaxi service
Wing Drone delivery Autonomous delivery network with Walmart and other retailers
Calico Aging research Aging biology programs and clinical candidates
X, the moonshot factory Incubator Tapestry (electric grid), Bellwether (geospatial risk), and spinouts such as Taara, Chorus, and Anori
Verily Health (former Other Bet) Precision health, now independent with Alphabet as a minority investor Verily Pre platform, Verily Me, Violet AI companion

History

Google's AI history began well before the generative AI boom. The Google Brain project started in 2011, the company acquired DeepMind in 2014, it open-sourced TensorFlow in 2015, and it revealed its first TPU in 2016, the same year DeepMind's AlphaGo defeated Lee Sedol using deep reinforcement learning and Monte Carlo tree search. In 2017 Google researchers published "Attention Is All You Need," which introduced the Transformer architecture that underlies nearly every modern language model, and in 2018 Google released BERT, which soon improved Google Search. Google also pioneered federated learning on Android keyboards, and DeepMind followed AlphaGo with AlphaGo Zero, AlphaStar, and the generalist agent Gato.

The language model era arrived in stages. LaMDA was a conversational large language model that could generate text, translate languages, and answer questions, and PaLM was a larger and more capable model that could reason about complex questions and produce poems, code, scripts, emails, and letters. Imagen generated realistic images from text descriptions, and MusicLM generated music from text descriptions. When OpenAI's ChatGPT launched in late 2022, many observers argued that Google, despite inventing much of the underlying technology, had been slow to ship; Google responded with Bard in March 2023, the Google DeepMind merger in April 2023, and the first Gemini models in December 2023, and it renamed Bard to Gemini in February 2024.

Momentum built through 2024 and 2025. Gemini 1.5 introduced million-token context windows, AlphaFold 3 extended structure prediction to nearly all biomolecules, and the 2024 Nobel Prize in Chemistry recognized AlphaFold 2. Gemini 2.0 and 2.5 added native image generation and "thinking" (extended chain-of-thought reasoning), a Gemini Deep Think model reached gold-medal level at the 2025 International Mathematical Olympiad, and Gemini 3 Pro arrived in November 2025 together with the Nano Banana Pro image model and the Google Antigravity coding platform. By 2026 the debate had flipped from whether Google was "missing the boat" to whether its full-stack strategy would let it lead: the company shipped a rapid series of Gemini 3.x Flash models, launched the always-on Gemini Spark agent at I/O, and released its first Gemini 4 frontier models in the autumn.

Gemini Foundation Models

Gemini is a family of natively multimodal foundation models that accept text, images, audio, video, and PDFs, and the family now spans a frontier tier (Gemini 4), a high-volume "workhorse" Flash tier, lightweight Flash-Lite models, and the on-device Gemini Nano. Google has not published full architectural details for recent models, but its model cards describe sparse mixture-of-experts Transformers with context windows of about one million tokens and adjustable "thinking levels" that trade latency for deeper reasoning. Developers reach these models through the Gemini API in Google AI Studio and through Google Cloud, while consumers meet them in the Gemini app, Search, Workspace, Android, and Chrome. A key 2026 pattern is that Google shipped the Flash tier on a fast cadence while its flagship tier changed names: the Gemini 3.5 Pro model previewed at I/O never shipped broadly, and the company moved directly to Gemini 4.

Gemini 4 Argon

Gemini 4 Argon, announced by Koray Kavukcuoglu on September 30, 2026, was Google's first new frontier model since Gemini 3.1 Pro and introduced a naming scheme based on chemical elements. Google positioned Argon for long-horizon work in real-world software engineering, enterprise knowledge work such as legal and financial analysis, and cybersecurity defense, and it highlighted an output limit of one million tokens, which lets a single agent run produce very large codebases or documents. Google reported a state-of-the-art 77.9 percent on the DeepSWE v1.1 software engineering benchmark and first place on Zapier's AutomationBench, and independent analysts noted Argon's unusually low hallucination rate, meaning it tends to admit uncertainty rather than guess. Access was staged: Argon went first to vetted cyber defenders in the Fairwind Program, then became generally available on October 13, 2026 through the paid Gemini API, Google AI Studio, Antigravity for AI Ultra subscribers, and the Gemini app and AI Mode for AI Pro and Ultra subscribers, at an introductory price of $2 per million input tokens and $10 per million output tokens.

Gemini 4 Carbon

Gemini 4 Carbon, announced on October 28, 2026, is Google's strongest model to date and is aimed at cybersecurity and very long agentic tasks. Google reported that Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, narrowly ahead of Anthropic's Claude Opus 5.5 at 58 and OpenAI's GPT-6 Astra at 57, and that it leads security benchmarks for finding vulnerabilities (88.2 percent on SEC-bench Pro), exploiting them in controlled tests (54.8 percent on ExploitGym), and patching them (71.0 percent on CWE-bench). Google says internal teams have run Carbon for days at a time on multi-step engineering tasks, such as tracing subtle bugs to their root cause across large codebases, and that the model is designed to ask a human for guidance when a task is ambiguous instead of forcing a path forward; the company also reported strong resistance to indirect prompt injection in third-party red-team testing. Because these capabilities are dual-use, Carbon is initially limited to a subset of cyber defenders and will expand to trusted partners through the Fairwind Program, with no date yet for general developer or consumer access, and some independent summaries of benchmark data report a higher hallucination rate than Argon. Carbon launched at the same introductory price as Argon, $2 per million input tokens and $10 per million output tokens with cached input discounted by 95 percent.

Gemini 3.x and the Flash Workhorse Line

Most everyday Gemini traffic runs on the Flash tier, which Google updated at a remarkable pace in 2026. Gemini 3 Pro (November 2025) and Gemini 3 Flash (December 2025) were followed by Gemini 3 Deep Think and Gemini 3.1 Pro in early 2026, and at I/O on May 19, 2026 Google launched Gemini 3.5 Flash as the new default in the Gemini app and AI Mode. When the promised Gemini 3.5 Pro slipped, Google instead released Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a security-focused Gemini 3.5 Flash Cyber in July, then Gemini 3.7 Flash on August 13 (about three weeks later) at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens. Gemini 3.8 Flash followed on September 2, 2026 as Google's most capable Flash model for long-horizon software engineering, autonomous agents, and enterprise workflows; it became the default model for Antigravity's managed agents and reached Google AI Pro and Ultra subscribers in the Gemini app, AI Mode, and Google Sheets. Critics, including reviewers at Ars Technica, questioned whether three-week release cycles reflected real progress or the appearance of it, while developers generally welcomed the falling prices.

  • Gemini Nano: the on-device model that runs on Pixel phones and other Android devices through Android's AICore service, powering private features such as call summaries, scam detection, and proactive suggestions without sending data to the cloud.
  • Gemini Live and Live API: low-latency audio-to-audio models for real-time spoken conversation with camera and screen context; a Gemini 3.8 Live model followed in September 2026.
  • Gemini text-to-speech: controllable speech generation models, including Gemini 3.1 Flash TTS with more than 200 audio tags and later Gemini 3.8 Flash TTS variants for narration, dubbing, and voice agents.
  • Gemini Embedding: multimodal embedding models (Gemini Embedding 2 in 2026) used for semantic search and retrieval-augmented generation.
  • Cyber variants: restricted models such as Gemini 3.5 Flash Cyber and Gemini 3.8 Flash Cyber, offered to vetted defenders through the Fairwind Program rather than as public API products.
  • Deep Think: an extended-reasoning mode that explores many solution paths in parallel and reached gold-medal performance at the 2025 International Mathematical Olympiad.

Generative Media Models

Google DeepMind builds a full suite of generative media models, and in 2026 it began unifying them under the Gemini brand. Image generation and editing now runs on the "Nano Banana" family, which began as the codename for Gemini 2.5 Flash Image in August 2025: Nano Banana Pro (Gemini 3 Pro Image) arrived in November 2025, Nano Banana 2 (Gemini 3.1 Flash Image) became the default in the Gemini app on February 26, 2026, and lighter and newer versions followed during the year. Google reported more than 50 billion images generated with these models by I/O 2026, and they power image features in Gemini, Search, Google Photos, Flow, and the new Google Pics app; the older Imagen models remain available to developers. For music, the Lyria models (including Lyria 3 and Lyria RealTime) generate instrumental and vocal tracks and power the Flow Music experiment in Google Labs, extending the earlier Magenta research project.

Video generation changed most. Google's Veo models (Veo 3 in 2025 added native sound effects, dialogue, and music, and Veo 3.1 added reference images and scene extension) remain the specialist choice for cinematic quality. At I/O 2026, Google DeepMind introduced Gemini Omni, a model that accepts any combination of text, images, audio, and video and outputs video, which lets users edit a clip conversationally over several turns instead of re-prompting from scratch. Google's developer documentation now recommends Gemini Omni Flash as the default video model and keeps Veo 3.1 for features such as frame-specific generation, and Omni also appears in the Gemini app, Google Flow, YouTube Shorts Remix, YouTube Create, and Google Vids. Every image, video, and audio clip from these models carries an invisible SynthID watermark, a safeguard discussed in the safety section below.

World Models and Simulation

World models are generative systems that simulate an environment frame by frame and respond to a user's or agent's actions, and Google DeepMind treats them as a stepping stone toward general intelligence. Genie 3, announced in August 2025, was the first of its models to support real-time interaction at about 20 to 24 frames per second, and it could keep a generated world consistent for a few minutes; DeepMind used it to train and test its SIMA game-playing agent, a form of simulated environment learning. In January 2026 Google opened Project Genie, an experimental prototype that lets Google AI Ultra subscribers create and explore their own worlds, and on October 27, 2026 it announced Genie 4 as a research preview for trusted testers. Genie 4 adds native audio (ambient sound, music, and interactive dialogue), more realistic physics such as collisions and aerodynamics, and better human motion, and it is coming to Project Genie and to early-access partners in the Gemini API; reviewers noted, however, that its worlds still stay coherent for only a couple of minutes. Uses range from game prototyping and film previsualization to robot training, where mistakes in simulation cost nothing.

Gemma Open Models

Gemma is Google's family of open models built from the same research as Gemini, intended to run on developers' own hardware, from phones to data center servers. Earlier generations were downloaded more than 400 million times and spawned over 100,000 community variants on platforms such as Hugging Face, and on April 2, 2026 Google released Gemma 4 under the permissive Apache 2.0 license, a notable change from the custom terms that governed earlier versions. Gemma 4 comes in four sizes: "effective" 2B and 4B models for phones and edge devices, a 26B mixture-of-experts model that activates only about 3.8 billion parameters per token for speed, and a 31B dense model designed for quality and fine-tuning. The larger models support context windows of up to 256,000 tokens and more than 140 languages, and Gemma even runs in orbit on the Project Suncatcher test satellite launched in October 2026.

  • MedGemma and TxGemma: open medical and therapeutic models (see the healthcare section).
  • ShieldGemma: safety classifiers for filtering harmful inputs and outputs.
  • PaliGemma: vision-language models for captioning, detection, and visual question answering.
  • EmbeddingGemma: compact on-device embedding models.
  • DolphinGemma: a model trained on dolphin vocalizations with the Wild Dolphin Project, an example of AI for animal communication research.
  • VaultGemma: a model trained with differential privacy to limit memorization of training data.
  • CodeGemma, T5Gemma, and DataGemma: variants for code, encoder-decoder tasks, and grounding answers in Google's Data Commons statistics.

The Agentic Era: Gemini Spark, Skills, and AI Studio

Google has shifted its AI strategy toward an "agentic" ecosystem in which Gemini moves beyond chat to become a proactive assistant that can plan, reason, and take action across digital environments. The Gemini app is the center of that strategy: at I/O 2026 Google gave it a full redesign called "Neural Expressive," made Gemini 3.5 Flash the default model, added a Daily Brief agent, and introduced Gemini Spark. Subscriptions are tiered: free users get the basic app, Google AI Plus and AI Pro add more capable models and features, and Google AI Ultra, which Google cut to $100 per month at I/O 2026, includes the most advanced agents, video tools, and early access programs. The app's capabilities are easiest to scan as a list.

  • Gemini Spark: an always-on, 24/7 personal AI agent built on Gemini 3.5 and the Antigravity agent harness. Spark runs on dedicated cloud virtual machines rather than on the user's device, so it keeps working when a phone is locked or a laptop is off; it can manage email, calendars, Docs, Sheets, and Slides, browse the web through Chrome, and even receive tasks through a dedicated Gmail address. It rolled out to testers and then U.S. AI Ultra subscribers in May 2026, gained access to local files through the Gemini app for macOS at the end of June, moved to Gemini 3.7 Flash in August, and was available to AI Pro and Ultra subscribers in more than 160 countries by late summer. Google requires explicit permission before Spark accesses a new service or takes an irreversible action.
  • Gemini Gems and Skills: Gems, introduced in 2024, are customizable versions of Gemini that act as specialized experts (a coding assistant, career coach, or writing editor) built from saved instructions and knowledge files. Google is replacing them with Skills, reusable Markdown-based instruction sets that can be combined for complex tasks; Skills began rolling out in Workspace on October 5 and in the Gemini app on October 13, 2026, existing Gems begin migrating automatically on November 17, 2026, and business and education accounts keep Gems until March 1 and June 1, 2027 respectively.
  • Google AI Studio: the primary developer playground for prototyping and testing Gemini models. It provides direct access to model parameters and "vibe coding" tools, allowing developers to build, tune, and deploy agentic workflows and custom applications before moving to production on Google Cloud.
  • Deep Research: an agent that searches the web and a user's files, then writes cited, multi-page reports.
  • Canvas: a workspace for drafting documents, code, web apps, and interactive visualizations alongside the chat.
  • Gemini Live: real-time spoken conversation with camera and screen sharing, built into the redesigned app.
  • Personal Intelligence and memory: optional context from Gmail, Photos, YouTube, and past chats so answers reflect a user's own history, governed by privacy controls.
  • Notebooks: research notebooks shared between the Gemini app and the standalone Google Gemini Notebook product.
  • Learning and creative modes: Guided Learning for step-by-step tutoring, Storybook for illustrated stories, and image and video creation with Nano Banana and Gemini Omni.
  • Multimodal capabilities: modern Gemini models process and reason across text, images, video, audio, and code, enabling tasks from cinematic video generation with Gemini Omni to automated infrastructure diagnostics.

Consumer Products and Platforms

Google's consumer reach is its largest structural advantage, because a model improvement can reach billions of people the day it ships. In 2026 Google pushed Gemini into every major surface at once (Search, Chrome, Android, Pixel, Home, Workspace, YouTube, Photos, and Maps), and it increasingly frames those surfaces as places where agents act on a user's behalf. The subsections below describe the main consumer offerings and how they changed during the year.

Search: AI Overviews, AI Mode, and Search Agents

Search remains Google's most important AI product, and Pichai described its 2026 overhaul as the deepest change in the product's history. AI Overviews, the generated summaries at the top of results, surpassed 2.5 billion monthly users, and AI Mode, a conversational search experience that breaks a question into sub-queries and synthesizes the answers, reached one billion monthly users within a year of launch. At I/O 2026 Google made Gemini 3.5 Flash the default model for both, began merging AI Overviews and AI Mode into a single interface with follow-up questions, and redesigned the search box for multimodal input for the first time in about 25 years. It also announced Search Agents (initially called information agents) that monitor the web in the background for things like apartment listings, ticket availability, and price drops, along with agentic coding and custom mini-apps generated directly in results; Google Lens, Circle to Search, and Search Live (real-time camera conversations) round out the visual search tools. Publishers and regulators continue to question how AI answers affect traffic to the websites they summarize, a debate covered in the challenges section.

Chrome and the Agentic Web

Chrome has become both an AI product and an agent platform. Gemini in Chrome can summarize and compare open tabs, answer questions about a page, and complete multi-step web tasks, and Google said Gemini Spark would gain the ability to work through Chrome later in 2026. For developers, Google proposed WebMCP, an open web standard that would let websites expose structured tools to browser-based agents, and it added Chrome DevTools features and "Modern Web Guidance" for coding agents. Because agents that read web pages are exposed to hidden instructions, Chrome's agent features include defenses against prompt injection attacks, and Google is adding C2PA content credentials and SynthID verification to Chrome and Search so users can check whether an image was generated by AI.

Android, Pixel, Wearables, and Android XR

Android 17 introduced "Gemini Intelligence," a system-level layer that lets Gemini take actions across apps, and Android Halo, an indicator at the top of the screen that shows what AI agents are doing on the device. The Pixel 11 family, announced in August 2026, runs on the Google Tensor G6 chip, which Google says has 50 percent more TPU compute than its predecessor and runs the latest Gemini Nano model up to 3.5 times faster while using up to 3.5 times less energy; on-device features include Magic Cue suggestions, Magic Capture, Camera Coach, voice translation during calls, and a Rambler voice input mode. Wear OS watches, including Pixel Watch 5, carry Gemini as well, and Pixel devices remain Google's showcase for private, on-device AI. Gemini also reaches cars through Android Auto, televisions through Google TV, and laptops through ChromeOS and the Googlebook laptops previewed at I/O 2026.

Extended reality is the newest hardware frontier. Android XR, built with Samsung and Qualcomm, launched on Samsung's Galaxy XR headset in October 2025, and at I/O 2026 Google previewed "intelligent eyewear": audio glasses from Gentle Monster and Warby Parker that speak Gemini responses privately into the wearer's ear are due in fall 2026 with iPhone support, while display glasses that project information into the wearer's view will follow. Partners such as Xreal are building tethered glasses like Project Aura, extending Gemini into immersive computing. The approach contrasts with Google Glass a decade earlier by centering the experience on conversational AI rather than on a camera.

Google's models also reach rival platforms. In January 2026 Apple and Google announced a multi-year, non-exclusive partnership under which the next generation of Apple Foundation Models, which power Siri and Apple Intelligence, will be based on Gemini models and Google's cloud technology, and Google Cloud CEO Thomas Kurian later described Google as Apple's preferred cloud provider for that work. The deal means Gemini technology now underpins AI features on both major mobile operating systems, a concentration that critics, including xAI founder Elon Musk, quickly questioned.

Google Home and Nest

Google is replacing Google Assistant on smart speakers and displays with Gemini for Home, a voice assistant that understands more natural requests, follow-up questions, and complex automations. Early access began in the United States in October 2025 and expanded to other countries, including Australia and Japan, in 2026. Advanced features such as Gemini Live conversations, natural-language searches of Nest camera history, and AI-generated camera alerts and summaries require Google Home Premium, the subscription that replaced Nest Aware, with Standard and Advanced plans starting at about $10 and $20 per month. Google launched the Gemini-native Google Home Speaker on June 25, 2026 at $99.99, and on July 23, 2026 extended Gemini Live to first-generation Google Home Mini and Nest Hub devices, making the home a major new surface for conversational AI and the Internet of Things.

Workspace and Productivity

Gemini is built into Gmail, Docs, Sheets, Slides, Meet, Drive, and Chat for Workspace business and consumer subscribers, continuing the effort that began as Duet AI in 2023. It drafts and summarizes email, analyzes spreadsheets and generates formulas in Google Sheets, builds slide decks, takes meeting notes, and translates speech live in Meet. At I/O 2026 Google announced Gmail Live, which lets users ask spoken questions about their inbox, and Docs Live for voice-driven editing, while Google Vids generates narrated videos (now with Gemini Omni) and Google Pics, a new image app built on Nano Banana, supports object-level edits, editing and translating text inside images, and shared collaboration; Pics reached broader availability in September 2026. Workspace Studio lets teams build no-code "flows" that chain Gemini steps across apps, and Workspace is the first place Gemini Skills rolled out, positioning Gemini as a personal productivity layer rather than a separate tool.

Gemini Notebook (formerly NotebookLM)

Google Gemini Notebook is Google's source-grounded research assistant: users upload documents, websites, videos, and notes, and the tool answers questions with citations to those sources, a practical application of retrieval-augmented generation. The product began as Project Tailwind at I/O 2023, launched as NotebookLM, and became one of Google's breakout AI hits thanks to Audio Overviews, which turn sources into podcast-style conversations, and later Video Overviews and study tools. On July 16, 2026, Google renamed it Gemini Notebook while keeping it a standalone product at the same web address, and it reported more than 30 million users. The update added a secure cloud computer that writes and runs code to analyze uploaded sources, and notebooks now sync with the Gemini app and are coming to Google Search, making Gemini Notebook a shared research layer across Google's ecosystem.

YouTube, Photos, Maps, Translate, and Shopping

Several other consumer products use AI for specific everyday tasks. YouTube applies recommendation models at enormous scale and in 2026 added Ask YouTube, a conversational way to search inside videos, along with Gemini Omni video generation in Shorts Remix and the YouTube Create app, automatic dubbing, and likeness detection that helps creators find unauthorized AI copies of their faces. Google Photos offers Ask Photos for natural-language searches of a personal library and Nano Banana editing tools, Google Maps uses Gemini for conversational navigation and place summaries, and Google Translate offers live speech translation and language practice built on Gemini's translation and speech recognition capabilities.

Shopping became an agentic category in 2026. Universal Cart, announced at I/O, is a Gemini-powered cart that works across the Gemini app, YouTube, and Gmail and tracks deals, price history, price drops, and stock alerts, while AI Mode supports virtual try-on and agentic checkout within a budget the user sets. Underneath these features sit open protocols that Google introduced with partners, such as the Agent Payments Protocol (AP2) for agent-initiated payments and the Universal Commerce Protocol, which aim to let agents from any company transact with merchants safely, a key piece of the interoperability layer for agent commerce.

Google Labs and Use-Case-Specific Apps

Google Labs is the company's public test bed for experimental AI products, and many of Google's most distinctive use-case-specific apps start there before graduating into full products, as NotebookLM, Flow, and Jules did. Labs experiments are typically free or bundled with AI subscriptions, launch first in a few countries, and are shut down or merged when they do not catch on; Whisk, an image remixing tool, was folded into Flow on April 30, 2026, and the Mixboard concept board closed on September 28, 2026. The table below lists the main Labs and use-case apps as of October 2026, from creative tools to no-code builders and science prototypes.

App Use case Notes (as of October 2026)
Flow AI filmmaking: generating, editing, and sequencing video scenes with consistent characters Runs Veo and Gemini Omni; absorbed Whisk's image remixing in April 2026
Flow Music Creating original songs with Lyria models Labs experiment
Stitch Turning prompts, sketches, wireframes, and screenshots into editable UI designs and front-end code Added "vibe design" in March 2026 and open-sourced its DESIGN.md specification for keeping agents on-brand
Opal Building shareable AI mini-apps by chaining prompts, models, and tools in plain language Expanded to more than 160 countries in November 2025; also available inside the Gemini web app
Pomelli Generating on-brand marketing content and product photography for small businesses Expanded to more than 170 countries by March 2026; relevant to marketing automation
Jules Asynchronous coding agent that fixes bugs, writes tests, and builds features in a GitHub repository Out of beta since 2025; complements Antigravity
Project Genie Creating and exploring interactive generated worlds Genie 3 today for AI Ultra; Genie 4 waitlist open
Gemini for Science experiments Hypothesis Generation (built on Co-Scientist) and Computational Discovery (built on AlphaEvolve and ERA) Launched at I/O 2026 at labs.google/science
Disco (GenTabs) Remixing open browser tabs into custom, task-specific web apps Waitlisted browser experiment since December 2025
Learn Your Way Personalizing textbook material into interactive lessons and quizzes Education research experiment; see Education
Dreambeans Daily AI-generated stories personalized from a user's Google apps Launched June 2026
Stax Evaluating models and prompts against custom criteria Developer toolkit for model evaluation
Little Language Lessons Bite-sized, situational language learning Free Labs experiment
Doppl Virtual try-on of outfits using a photo of the user Launched 2025; related to fashion commerce
Mixboard AI concept and mood boards using Nano Banana Shut down September 28, 2026

Developer Platforms and Tools

Google offers developers a layered toolkit: hosted model access through the Gemini API, agent-building environments such as Google Antigravity, open frameworks such as TensorFlow and JAX, and on-device runtimes for phones and embedded hardware. The direction in 2026 was consolidation around agents, with Google retiring several older tools and unifying its coding products under the Antigravity brand. For side-by-side comparisons with rival tools, see the AI Coding Agent Comparison and Agent Development Framework Comparison pages.

Google AI Studio and the Gemini API

Google AI Studio is the free, browser-based entry point for prototyping with Gemini, Gemma, Nano Banana, Veo, Gemini Omni, Lyria, and Gemini Robotics ER models, and it can generate a working web app from a description and deploy it with one click. The Gemini API behind it offers text, image, video, audio, live, and embedding endpoints, a newer Interactions API for multi-turn sessions, built-in tools such as Google Search grounding, URL context, code execution, and computer use, and batch, flex, and priority consumption tiers. At I/O 2026 Google added Managed Agents to the Gemini API so developers can deploy persistent agents without running their own infrastructure, and those agents run on the same harness as Antigravity. Applications typically move from AI Studio to Google Cloud when they need enterprise governance, data residency, or service-level agreements.

Google Antigravity

Google Antigravity launched in November 2025 alongside Gemini 3 as an "agent-first" development environment in which AI agents plan, write, test, and verify code while the developer supervises. At I/O 2026 Google turned it into a full platform with Antigravity 2.0: a desktop app that orchestrates multiple agents and subagents in parallel, an Antigravity CLI rewritten in Go that replaces the Gemini CLI, an SDK (initially a Python research preview) that exposes the agent harness for custom agents, and the original IDE on a slower track. Google also retired the Gemini Code Assist IDE extensions for most individual users in favor of Antigravity, while enterprise customers kept access. The same harness powers Gemini Spark and the Gemini API's Managed Agents, and Antigravity gained Science Skills, a bundle that connects agents to more than 30 life science tools and databases.

Coding, Mobile, and Data Science Tools

Beyond Antigravity, Google supports developers across the software lifecycle. Jules works asynchronously on GitHub repositories, Firebase offers AI features for app backends (Firebase AI Logic) and integrates with AI Studio, and Android Studio includes Gemini for code generation and, with Gemma 4, local agentic coding for Android development. For data work, Colaboratory provides free hosted Jupyter notebooks with a Gemini-powered Data Science Agent, Kaggle hosts competitions, datasets, and model benchmarks, and BigQuery ML lets analysts train models with SQL.

Agent Protocols and Frameworks

Google has invested heavily in open standards so that agents built by different companies can cooperate. The Agent Development Kit (ADK) is an open-source framework for building multi-agent systems in Python and other languages; the Agent2Agent (A2A) protocol, contributed to the Linux Foundation in 2025, lets agents from different vendors discover each other and delegate tasks; and Google supports Anthropic's Model Context Protocol (MCP) for connecting agents to tools and data. Commerce-specific protocols (AP2 and the Universal Commerce Protocol) and the proposed WebMCP browser standard extend the same idea to payments and websites. These efforts aim to make agents interoperable rather than locked into one platform, a theme explored on the AI Marketplace & Toolkit/Model Interoperability page.

Open-Source Frameworks and Edge AI

Google's open-source machine learning stack remains widely used in research and industry. TensorFlow and its tools (TensorBoard, TensorFlow Extended (TFX), TensorFlow.js, TensorFlow Hub, and TensorFlow Serving) are joined by JAX, a high-performance numerical library built for differentiable programming that Google DeepMind uses to train Gemini, and by Keras, now a multi-backend API that runs on JAX, TensorFlow, and PyTorch. The XLA compiler and the OpenXLA project optimize models for TPUs, GPUs, and CPUs, and Kubeflow Pipelines orchestrates machine learning workflows on Kubernetes.

For on-device AI, LiteRT (the new name for TensorFlow Lite), MediaPipe, and the Google AI Edge toolkit run models on phones, browsers, and microcontrollers, while Google Coral supplies edge TPU hardware and ML Kit packages common vision and language tasks for mobile apps. Educational and research tools include Teachable Machine for training simple image, sound, and pose models without code, Google Facets for dataset visualization, the Language Interpretability Tool (LIT) and other People + AI Research (PAIR) projects for interpretability, Dopamine for reinforcement learning research, and DialogFlow for chatbots and voice bots. Separately from its AI work, Google engineers also maintain Carbon, an experimental open-source programming language introduced in July 2022 as a possible successor to C++; it remains experimental and is unrelated to the Gemini 4 Carbon model.

Google Cloud AI

Google Cloud sells Google's AI to enterprises, governments, and startups, and its 2026 strategy centers on what Thomas Kurian called "the agentic enterprise." At Google Cloud Next in Las Vegas on April 22, 2026, Google said its models were processing more than 16 billion tokens per minute through its APIs, up from 10 billion the previous quarter, and that most Google Cloud customers now use its AI products. The main offerings fall into four groups: the Gemini Enterprise app and agent platform, applied AI and data services, AI infrastructure, and AI-driven security.

Gemini Enterprise and the Gemini Enterprise Agent Platform

Gemini Enterprise, introduced in October 2025 as the successor to Agentspace, is a single workplace app where employees chat with Gemini, search company data, and use or build agents. At Next 2026 Google expanded it with Agent Designer for no-code, schedule-based or trigger-based agents, an Inbox that tracks long-running agent activity, shared Projects connected to Drive and Gemini Notebook, Skills for repeatable tasks, and a Canvas that edits documents and slides and exports them in Microsoft 365 formats. Underneath, Google rebranded Vertex AI as the Gemini Enterprise Agent Platform, a unified environment to build, scale, govern, and optimize agents that combines model selection, tuning, agent orchestration, identity, and audit controls. Its Model Garden offers more than 200 models, including Gemini, Gemma, Imagen, Veo, Lyria, and third-party models such as Anthropic's Claude family, and early customers included KPMG, which reported deploying more than 100 agents in a month; see also the Visual and Low-Code Agent Builder Comparison.

Applied AI and Data Services

Google Cloud also sells task-specific AI services that predate the agent era and remain heavily used. These include Document AI for extracting structured data from forms and contracts (document understanding), Vision AI and Video Intelligence for image and video analysis, Speech-to-Text and Text-to-Speech built on the Chirp speech models (speech synthesis), Translation AI, the Natural Language API, and the Customer Engagement Suite for contact centers. On the data side, BigQuery, Looker, and AlloyDB embed Gemini for natural-language queries, and the "Agentic Data Cloud" announced at Next 2026 connects agents directly to enterprise databases, alongside new database options such as Spanner Omni, which entered preview at the event. Older offerings listed in this page's navigation, such as the AI Platform, ML Engine, Prediction API, and Google AutoML, have largely been folded into these newer services.

AI Infrastructure: TPUs and AI Hypercomputer

Google's custom TPUs are the foundation of its AI economics, and they are now a product in their own right. After earlier generations such as Trillium (v6e) and Ironwood (v7, 2025), Google announced its eighth generation at Next 2026 as two specialized chips, TPU 8t for training and TPU 8i for inference, with general availability planned for late 2026. Google sells TPU capacity through its AI Hypercomputer architecture, which also includes NVIDIA GPUs, Axion Arm-based CPUs, and liquid-cooled data centers, and major outside model developers, notably Anthropic, run frontier models on TPUs; Apple's Gemini-based foundation models are also being developed with Google Cloud. Google's chip design itself uses AI, since DeepMind's AlphaChip has helped lay out several TPU generations.

Security: Wiz, Mandiant, AI Threat Defense, and the Fairwind Program

Security has become one of Google's most important AI businesses, and also one of its most sensitive, because the same models that find vulnerabilities could help attackers exploit them. Google completed its $32 billion acquisition of the cloud security company Wiz on March 11, 2026, its largest acquisition ever, and combined it with Mandiant incident response, Google Threat Intelligence, and Google Security Operations, which gained new defensive agents at Next 2026. On May 28, 2026, Google Cloud launched AI Threat Defense, a pipeline that aims to find, assess, fix, and monitor vulnerabilities automatically, and CodeMender, a DeepMind agent announced in October 2025 that had already upstreamed dozens of security fixes to open-source projects, entered preview for Google Cloud customers in July 2026.

The Fairwind Program, launched on September 2, 2026, is Google's answer to the dual-use problem. It gives a vetted group of more than 650 governments, critical-infrastructure operators, and security partners exclusive access to Google's most capable cyber models, starting with Gemini 3.8 Flash Cyber and extending to Gemini 4 Argon and Gemini 4 Carbon, combined with the CodeMender harness so defenders can find, verify, and patch vulnerabilities at scale; other customers can still use CodeMender with publicly available models. Participants must agree to operational controls such as restricting access to security teams and using multi-factor authentication. Related research projects include Big Sleep, an agent that has discovered previously unknown vulnerabilities in widely used software, and Google's Secure AI Framework (SAIF), which guides customers on defending AI systems; see also Cybersecurity and Cybersecurity Companies/Products.

Science: Google DeepMind and Google Research

Google's scientific AI program is one of the clearest differences between it and most AI companies, and it has produced peer-reviewed results across biology, mathematics, physics, and weather. The underlying idea, which Demis Hassabis summarized in his 2024 Nobel lecture as using AI to accelerate scientific discovery, is that the same learning methods that power chatbots can search vast spaces of molecules, proofs, and algorithms far faster than humans. In 2026 the emphasis shifted from bespoke systems for single problems toward general Gemini-based research agents, a change that also contributed to the breakup of the original AlphaFold team (covered in the next section).

Gemini for Science and AI Co-Scientist

At I/O 2026 Google announced Gemini for Science, a collection of tools and experiments meant to speed up core steps of the scientific method. Its Hypothesis Generation prototype is built on Co-Scientist, a multi-agent Gemini system that holds an "idea tournament" in which agents generate, debate, and rank hypotheses with citations to the literature, and its Computational Discovery prototype combines AlphaEvolve with Empirical Research Assistance (ERA) to generate and test thousands of code variants for computational experiments. Papers describing Co-Scientist and ERA were published in Nature the same week, and Google said organizations including Daiichi Sankyo, Bayer Crop Science, and U.S. National Laboratories participating in the Department of Energy's Genesis Mission were already using Co-Scientist, while companies such as BASF and Klarna were testing AlphaEvolve in private preview. A Science Skills bundle in Antigravity connects agents to scientific databases and tools, and reviewers such as Wharton's Ethan Mollick observed that the early tools lean toward biology rather than the social sciences.

Mathematics, Algorithms, and Computing

AlphaEvolve, introduced in May 2025, is an evolutionary coding agent that pairs Gemini models with automated evaluators to discover better algorithms, an application of evolutionary computation at the level of whole programs. It found faster procedures for some matrix multiplication problems, improved Google's data center scheduling and chip designs, and recovered measurable compute across Google's fleet. Earlier systems in the same lineage include AlphaTensor for matrix multiplication, AlphaDev for sorting routines, and AlphaChip for chip floorplanning, while AlphaProof and AlphaGeometry reached silver-medal level at the 2024 International Mathematical Olympiad and a Gemini Deep Think model reached gold in 2025. Google DeepMind leaders have also said that AI agents now run parts of the company's own model-training research by conducting experiments, analyzing results, and proposing hypotheses under human supervision.

Quantum AI

Google Quantum AI builds superconducting quantum processors and uses machine learning to make them work. Its Willow chip, revealed in December 2024 with 105 qubits, was the first to show that adding more qubits to an error-correcting code reduced errors exponentially, a long-sought milestone known as operating "below threshold," and DeepMind's AlphaQubit decoder uses neural networks to identify quantum errors more accurately than earlier methods. In October 2025 Google reported its "Quantum Echoes" algorithm as a verifiable demonstration of quantum advantage on a physics problem, and in 2026 it began accepting research proposals for outside access to Willow. Quantum computing remains a research effort rather than a commercial AI product, but Google presents it as a long-term complement to AI for chemistry and materials simulation.

Materials, Fusion, Physics, and the Humanities

DeepMind's GNoME project used graph neural networks to predict the stability of millions of candidate crystal structures, adding hundreds of thousands of potentially stable inorganic materials to the known set, and the predictions are being tested by automated labs. In energy, Google DeepMind has applied reinforcement learning to control plasma in tokamak reactors and partners with Commonwealth Fusion Systems on fusion simulation and control, and some former AlphaFold researchers were reassigned to fusion work in 2026. Other projects include control methods that reduce noise in the LIGO gravitational-wave observatory, which supports astronomy and physics, and Aeneas, a model published in Nature in 2025 that helps historians restore, date, and place ancient Latin inscriptions, an example of AI in archaeology.

Protein Science, Biology, and Drug Discovery

Protein science is where Google's AI has had its deepest scientific impact, and it now spans structure prediction, protein design, genomics, biosecurity, and a dedicated drug discovery company. Proteins are chains of amino acids that fold into three-dimensional shapes, and a protein's shape largely determines what it does, so predicting and designing shapes is central to understanding disease and creating medicines. The subsections below cover each offering in turn; for background, see Protein Folding & Discovery, Molecular Artificial Intelligence (AI), Bioinformatics, and Life Sciences.

AlphaFold

AlphaFold 2 effectively solved the 50-year protein folding problem at the CASP14 assessment in 2020, predicting many structures with accuracy comparable to laboratory methods, and DeepMind and EMBL's European Bioinformatics Institute then released the AlphaFold Protein Structure Database, which holds more than 200 million predicted structures covering nearly every catalogued protein. AlphaFold 3, released with Isomorphic Labs in May 2024, extended predictions to complexes of proteins with DNA, RNA, small molecules, and ions, and is available to researchers through the free AlphaFold Server; some scientists criticized its initially restricted code release, which spurred open alternatives. Hassabis and Jumper shared half of the 2024 Nobel Prize in Chemistry for AlphaFold 2, and the system has become a standard tool in structural biology, drug discovery, and enzyme engineering.

The organization behind AlphaFold changed in 2026. In July the Financial Times reported, and Google DeepMind confirmed, that the dedicated AlphaFold team had been dissolved, with most of the original paper's authors reassigned over the previous year to Gemini, Isomorphic Labs, enzyme design, fusion, and genomics work. John Jumper, who had moved with colleague Jonas Adler to an internal AI coding team earlier in the year, announced in June that he was leaving for Anthropic, and Adler and Alexander Pritzel followed. Many scientists interviewed by Scientific American were not alarmed, noting that the AlphaFold database and server remain available and that structure prediction is now widely reproduced, but the episode illustrated Google's shift from standalone scientific moonshots toward Gemini-centered research.

Protein Design: AlphaProteo, AlphaProtein Novo, and SynthID Bio

Where AlphaFold predicts the shapes of existing proteins, Google's design systems create new ones. AlphaProteo, announced in September 2024, designs novel protein binders that latch onto target molecules; it was the first AI tool to design a successful binder for VEGF-A, a protein involved in cancer and diabetes complications, and it achieved 3 to 300 times better binding strength than the best existing methods on seven targets. On October 5, 2026, Google DeepMind released a preprint and code for AlphaProtein Novo, a diffusion-based pipeline that designs enzyme structures and sequences around a desired catalytic site; developed with Nobel laureate Frances Arnold's group, it reported state-of-the-art activity on benchmark reactions and enzymes that synthesize the pharmaceutical building block piperidine or degrade the plastic additive DEHP. These results come from preprints and still require independent replication.

Protein design raises biosecurity concerns, because tools that create useful proteins could in principle be misused. On September 30, 2026, Google DeepMind introduced SynthID Bio, which embeds an imperceptible, verifiable watermark directly into AI-designed protein sequences and predicted structures. In laboratory tests, the watermarked designs matched the performance and natural diversity of unwatermarked ones, which means provenance can be tracked without sacrificing function, extending the SynthID approach from media to chemistry and biology.

Genomics and Cell Biology

Google's genomics tools interpret DNA, the instructions that cells use to build proteins. AlphaMissense (2023) classifies the likely effect of tens of millions of single-letter "missense" mutations that change a protein, and AlphaGenome, published in Nature on January 28, 2026 with its model weights released to scientists, reads DNA sequences of up to one million letters and predicts how variants in the 98 percent of the genome that does not code for proteins affect gene regulation across 11 molecular processes. Researchers have already used AlphaGenome to explain leukemia-driving mutations, although critics such as Johns Hopkins computational biologist Steven Salzberg caution that its predictions depend heavily on imperfect training data, and even enthusiastic users report that its accuracy fades farther from the gene being studied. Older and related tools include DeepVariant and DeepSomatic for identifying genetic variants in sequencing data, and Cell2Sentence-Scale, a Gemma-based model developed with Yale that treats single-cell gene activity as text and generated a cancer-related hypothesis later supported by laboratory experiments.

Isomorphic Labs

Isomorphic Labs, spun out of DeepMind in 2021 and led by Demis Hassabis, aims to use AI to design medicines and, in its own words, to "solve all disease." In February 2026 it published a technical report on the Isomorphic Labs Drug Design Engine (IsoDDE), which it says improves on AlphaFold 3 at predicting how proteins bind to drug-like molecules and how strongly, and can identify new binding pockets from a protein sequence alone. In May 2026 it raised $2.1 billion in a Series B round led by Thrive Capital with Alphabet, GV, CapitalG, Temasek, MGX, and the UK's Sovereign AI Fund, following a $600 million round in 2025, and it has research partnerships with Eli Lilly and Novartis and sites in London, Cambridge, Massachusetts, and Lausanne. The company now expects its first AI-designed drugs to enter clinical trials by the end of 2026, later than the end-2025 target Hassabis had once suggested, and it will ultimately be judged on trial data rather than benchmarks; see Pharmaceuticals for the wider field.

Healthcare and Medicine

Google's health AI work spans clinical research systems, open models for medical developers, consumer wellness products, and two independent life-science companies that began inside Alphabet. Google consistently describes its clinical systems as research tools rather than medical devices, and most require further validation and regulatory review before routine use in care. For related topics see Healthcare, Pathology, and Wearables.

Clinical AI Research: AMIE, Med-Gemini, and MedGemma

AMIE (Articulate Medical Intelligence Explorer) is a research system from Google Research and Google DeepMind for diagnostic conversations, and it progressed rapidly in 2026. A Nature study published in June 2026 showed AMIE reasoning about longitudinal disease management in blinded comparisons with primary care physicians, an August 2026 study demonstrated expert-level performance in real-time video consultations with patient actors, and on October 8, 2026 The Lancet published a prospective feasibility study with Beth Israel Deaconess Medical Center in which real patients used AMIE from home to share their histories before urgent primary care visits. The investigators stressed that the study tested feasibility and safety, not whether AI improves health outcomes, and Google says more research is needed before clinical deployment; a nationwide IRB-approved study with Included Health is evaluating AI in virtual care.

For developers, Google evolved its earlier Med-PaLM and Med-Gemini research into open tools. MedGemma, first released in 2025 and built on Gemma, includes text and multimodal models for radiology reports, medical images, and clinical records, along with MedSigLIP for image classification, and its results were published in Nature Medicine; health systems and startups are fine-tuning it, from Singapore's Ministry of Health building a primary care model to Dawa Health integrating it into an offline app for maternal care in Zambia. MedGemma is part of Health AI Developer Foundations (HAI-DEF), a set of free open-weight models that also includes foundation models for dermatology, chest X-rays, digital pathology, and health acoustics, and TxGemma extends the approach to predicting properties of potential therapeutics. Google also runs screening programs that use AI to detect diabetic eye disease and support breast and lung cancer screening in partner health systems.

Consumer Health: Google Health App, Health Coach, and Fitbit

On May 19, 2026, Google replaced the Fitbit app with the Google Health app, consolidating fitness, wellness, and medical record data in one place and announcing the retirement of Google Fit. Its centerpiece, Google Health Coach, is a Gemini-powered assistant that creates adaptive fitness plans, answers questions about sleep, activity, and health data, and summarizes medical records, using a multi-agent design with separate components for conversation, data analysis, and coaching. The coach is part of Google Health Premium (formerly Fitbit Premium) at $9.99 per month or bundled with Google AI Pro and Ultra in many countries, and Google launched the screenless Fitbit Air tracker at $99 to pair with it while working to support data from Apple Health, Health Connect, and other wearables. Google says the coach is not a medical device and does not diagnose, a distinction regulators and medical groups are watching as rivals such as OpenAI, Microsoft, Samsung, Oura, and Whoop launch similar services.

Verily and Calico

Verily, which began at Google X and later became an Alphabet Other Bet, became an independent company in 2026. In March it raised $300 million in a round led by Series X Capital, converted from an LLC to a corporation named Verily Health Inc., and Alphabet became a significant minority investor without a controlling stake. Verily's AI-native precision health strategy centers on the Verily Pre data platform, the Verily Me consumer app with an AI companion called Violet, the Lightpath chronic-disease program, and Workbench, which supports more than 21,500 researchers in the NIH's All of Us Research Program; it also works with Samsung, Salesforce, and NVIDIA.

Calico (California Life Sciences), founded by Google in 2013 to study the biology of aging, uses machine learning alongside laboratory biology but has struggled to translate research into approved therapies. Its fosigotifator program failed to meet its primary endpoint in an amyotrophic lateral sclerosis trial in early 2025, and in November 2025 STAT News reported that AbbVie was ending the 11-year partnership that had funded much of Calico's drug development, after AbbVie had invested about $1.75 billion. Calico continues to advance programs such as ABBV-CLS-628, an antibody for polycystic kidney disease that received FDA fast-track and orphan drug designations, and it appointed a former AbbVie chemist to lead its drug discovery.

Climate, Carbon, and Sustainability

Google's climate work has two sides that are often discussed separately but are tightly linked. On one side, Google builds AI tools that help governments, airlines, cities, and utilities forecast hazards and cut emissions; Google's 2026 environmental report says nine of its solutions helped users and partners reduce an estimated 41 million tonnes of carbon dioxide equivalent in 2025, and the company has set a goal of helping reduce one gigaton of emissions per year by 2030. On the other side, the data centers that run those same AI models are driving Google's own electricity use and emissions upward, which complicates its goal of running on 24/7 carbon-free energy by 2030. Both sides are covered here; see also Environmental Science, Energy, and Power (Management).

Weather and Climate Modeling

Google DeepMind and Google Research have transformed weather forecasting with machine learning. GraphCast (2023) used a graph neural network to produce ten-day forecasts that rivaled leading physics-based models in about a minute, GenCast (2024) added probabilistic ensemble forecasts, and NeuralGCM combined machine learning with a traditional atmospheric model for climate research. These lines converged in WeatherNext, a family of models that can produce affordable forecasts up to 10 days ahead and a range of possible scenarios up to 15 days ahead; WeatherNext 2 generates high-resolution hourly forecasts in minutes, and a cyclone-specific model published in 2026 extended lead times for tropical storm tracks and intensity by about a day. The U.S. National Hurricane Center used WeatherNext guidance to anticipate Hurricane Melissa's landfall in Jamaica five days in advance, illustrating how AI forecasting has moved from research into operational warnings.

Crisis Resilience: Floods, Wildfires, Earthquakes, and Heat

Google's crisis resilience tools turn forecasts into alerts for the public and for first responders. Flood Hub, which grew from a 2018 pilot in India's Patna region, now covers about two billion people in more than 150 countries with riverine flood forecasts up to seven days ahead and urban flash-flood warnings up to 24 hours ahead; Google trained its flash-flood model on Groundsource, a dataset of 2.6 million historical flood records extracted by AI from 20 years of public reports, and it has open-sourced that dataset and its hydrology framework. For wildfires, Google tracks fire boundaries in dozens of countries and helped create FireSat with the Earth Fire Alliance and Muon Space, a satellite constellation designed to detect fires as small as 5 by 5 meters and refresh the global picture every 20 minutes; after a prototype launched in March 2025, the first three operational satellites launched on July 7, 2026. The Android Earthquake Alerts System uses accelerometers in Android phones as a global seismic network and delivered more than 11 million warnings during a 2026 earthquake sequence in Venezuela, and Google also provides heat, air quality, and post-disaster damage mapping tools.

Carbon and Emissions Tools

Several Google products target emissions directly, and they are among its most concrete "carbon" offerings. The most striking is Project Contrails: the thin clouds that form behind aircraft account for roughly a third of aviation's warming impact, and Google combines satellite imagery, weather data, and AI predictions so airlines can make small altitude changes that avoid contrail-forming regions. Trials with American Airlines reported contrail reductions of about 62 percent, the UK government and Google launched a 30-month trial called Operation Blue Skies in August 2026, and Cathay Pacific began Asia-Pacific trials in September 2026. Because only a small fraction of flights create most persistent contrails, the approach could cut warming at low cost, although it still needs regulatory acceptance and better atmospheric verification.

  • Environmental Insights Explorer (EIE): free emissions and climate data (building and transportation emissions, solar potential, tree canopy) for more than 40,000 cities and regions, used for greenhouse gas inventories and climate action plans.
  • Project Green Light: an AI tool that recommends traffic signal timing changes to city engineers, reducing stops by up to about 30 percent and intersection emissions by about 10 percent in early results, a practical smart city application.
  • Eco-friendly routing: Google Maps suggests routes with lower fuel or energy use, along with flight emissions estimates in Google Flights built on the open Travel Impact Model.
  • Solar API and Project Sunroof: rooftop solar potential estimates from aerial imagery and 3D models.
  • Carbon Footprint and the Carbon Sense suite: Google Cloud tools that report the emissions associated with a customer's cloud usage and recommend cleaner regions and ways to remove idle resources.
  • Tapestry: an X moonshot powered by Google Research, Google Cloud, and Google DeepMind that helps PJM, North America's largest grid operator, connect new power sources to the grid faster.
  • Gemini per-prompt footprint: in August 2025 Google published a methodology and measurements showing that the median Gemini Apps text prompt used about 0.24 watt-hours of electricity, emitted about 0.03 grams of carbon dioxide equivalent, and consumed about 0.26 milliliters of water, after 33-fold and 44-fold reductions in energy and carbon over 12 months.

Earth Observation and Geospatial AI

Google Earth Engine has hosted petabytes of satellite and environmental data for scientists since 2010, and Google now layers foundation models on top of it. Google Earth AI is a collection of geospatial models and datasets for floods, wildfires, population dynamics, and land use, used by governments, insurers, and public health researchers, and AlphaEarth Foundations (2025) compresses many satellite and climate data sources into a compact "embedding" for every 10-meter patch of land, released as a global Satellite Embedding dataset. Open Buildings maps structures across Africa, Asia, and Latin America where official records are sparse, and X's Bellwether project applies similar techniques to help insurers price risk in disaster-prone areas. Together these tools are building toward what Google calls geospatial reasoning, in which Gemini agents answer questions by combining many Earth models at once.

Nature and Biodiversity

Google also applies AI to ecology and conservation. Perch is a bioacoustics model that identifies birds and other animals in field recordings, an application of bird identification that helps scientists monitor ecosystems at scale, and SpeciesNet, an open model behind the Wildlife Insights platform, classifies animals in camera-trap images. DolphinGemma explores patterns in dolphin vocalizations, and Google supports partners such as Global Fishing Watch, which uses machine learning on satellite and vessel data to expose illegal fishing. These projects matter because biodiversity monitoring has traditionally depended on slow manual review of images and audio.

Google's Own Footprint and Project Suncatcher

AI has made Google's own energy picture harder. Google's 2026 environmental report described record data center electricity consumption driven by AI and acknowledged that its path to net-zero emissions has become more complex, and S&P Global's analysis found that the share of carbon-free electricity serving Google's data centers slipped from 66 percent in 2024 to 65 percent in 2025 even as the company signed record clean-power deals. Google has responded with large purchases of wind, solar, geothermal, and nuclear power (including approval for 150 megawatts of geothermal power in Nevada in February 2026 and support for restarting a nuclear plant in Iowa), it is the second-largest corporate buyer of carbon removal credits after Microsoft, and a senior executive said in May 2026 that the 24/7 carbon-free energy goal for 2030 remains intact. Google's long-running use of DeepMind AI to optimize data center cooling is another efficiency lever.

Project Suncatcher is Google's most speculative response to the energy problem. Announced in November 2025, it explores whether fleets of solar-powered satellites carrying TPUs could one day run AI workloads in orbit, where solar panels can generate up to eight times more energy than on the ground; see also space-based data centers. On October 1, 2026, Google and Planet launched a refrigerator-sized prototype carrying four TPUs on a SpaceX rideshare mission to test how the chips handle launch stress, radiation, and extreme temperatures, and the chips run a version of Gemma in short sessions because heat is hard to shed in a vacuum. Google plans to launch two more satellites in 2027 to test high-bandwidth laser links, but launch costs, cooling, maintenance, and the emissions from rockets themselves mean commercial deployment is far off.

Autonomous Systems and Robotics

Alphabet's "physical AI" offerings are its most visible use of AI outside screens. They include the largest commercial robotaxi service in the United States, a growing drone delivery network, foundation models for robots, and an industrial robotics software platform. See Robotics, Embodied AI, Transportation (Autonomous Vehicles), and Drones for broader context.

Waymo

Waymo, which began as Google's self-driving car project in 2009, operates fully driverless ride-hailing with its Waymo Driver system, which fuses lidar, radar, and camera data with machine learning models for perception, prediction, and planning. Its expansion accelerated sharply in 2026: Waymo launched paid driverless rides in Denver, San Diego, and Tampa on September 1, 2026, bringing its commercial service to 14 U.S. cities, and it reported more than 500,000 paid rides per week while targeting one million weekly rides by the end of the year. Waymo is also preparing service in London and announced plans for Toronto in October 2026, the same month it closed $5 billion in debt financing to fund expansion. Its research teams have explored large end-to-end multimodal driving models, and its safety record relative to human drivers is closely watched by regulators.

Wing

Wing, Alphabet's drone delivery company, flies autonomous aircraft that carry small packages from store parking lots to customers' yards, with human pilots monitoring many flights from a central operations hub. In January 2026 Wing and Walmart announced plans to add drone delivery at 150 more Walmart stores during the year, building toward more than 270 locations by 2027 and potentially serving about 40 million Americans in metro areas from Los Angeles to Miami. Wing also partners with other retailers and delivery platforms, and its growth depends on Federal Aviation Administration approvals for flights beyond visual line of sight.

Gemini Robotics and Intrinsic

Google DeepMind's Gemini Robotics models give robots a general-purpose "brain" built on Gemini, combining vision-language-action control with embodied reasoning about objects, space, and tasks. On July 30, 2026, DeepMind introduced the Gemini Robotics 2 family: Gemini Robotics 2, a vision-language-action model that controls whole humanoid bodies from walking to two-handed manipulation and lets multiple robots collaborate; Gemini Robotics ER 2, an embodied-reasoning model built on Gemini 3.5 Flash that plans, tracks progress, and calls tools, now in public preview in the Gemini API and AI Studio; and Gemini Robotics On-Device 2 for robots that must work without a network connection. The action models remain limited to partners such as Apptronik, Boston Dynamics, Franka Robotics, and Agile Robots, and reported success rates on tasks such as manipulating objects on shelves or using a dustpan vary widely, which shows how far general-purpose robots still have to go.

Intrinsic, an Alphabet company that builds software to make industrial robots easier to program, joined Google on February 25, 2026 after nearly five years as an independent Other Bet. Its Flowstate platform and vision models now connect directly to Gemini, Google DeepMind research, and Google Cloud, and it works with manufacturers such as Foxconn on AI-driven assembly. Together with the Boston Dynamics partnership announced at CES 2026, the move consolidates Google's robotics work around an "Android for robots" strategy in which Google supplies the intelligence and partners supply the hardware.

X, the Moonshot Factory

X, the moonshot factory, incubates ambitious technology projects and has produced several of Alphabet's AI companies, including Waymo, Verily, and Wing. Since 2025 it has spun out promising projects as independent startups funded by Series X Capital, a venture fund led by former Facebook chief financial officer Gideon Yu in which Alphabet is a minority investor. Recent beneficiaries include Taara (laser-based connectivity), Chorus (AI-powered supply chain orchestration), and Anori (AI for housing development), while projects such as Tapestry for the electric grid and Bellwether for geospatial risk remain closely tied to Google's climate work. For related ideas, see Moonshots and Supply Chain.

Education, Accessibility, and Learning

Education is one of Google's largest AI use cases because so many students already use Google Classroom, Chromebooks, and Search. LearnLM, a family of models fine-tuned on learning-science principles, now underlies Gemini's Guided Learning mode, which tutors step by step instead of simply giving answers, and Gemini for Education brings these tools to schools with administrative controls; Google has offered a free year of Google AI Pro to students aged 18 and older in countries including the United States, Japan, Indonesia, South Korea, and Brazil. Experiments such as Learn Your Way, Little Language Lessons, and Read Along explore personalized and language learning, and the Google Skills platform collects AI courses and certificates; see Education and AI Learning Guide for more.

Accessibility tools are some of Google's most practical AI applications. Project Euphonia and Project Relate help people with atypical speech be understood by speech recognition, Live Transcribe and Live Caption provide real-time captions, Lookout and TalkBack (now with Gemini image descriptions) help blind and low-vision users understand their surroundings and screens, and Guided Frame helps blind users take photos. Google researchers also study risks, including the cognitive effects of AI use on children, which shape how Google designs tutoring products.

Responsible AI, Safety, and Governance

Google's safety framework combines published principles, technical safeguards, and staged releases. Its AI Principles, first published in 2018, were revised in February 2025 to emphasize innovation, collaboration, and responsible development, and the revision drew criticism for removing explicit pledges not to build AI for weapons or surveillance. The Frontier Safety Framework defines "critical capability levels" in areas such as chemical, biological, radiological, and nuclear misuse, cyber offense, and harmful manipulation, and Google's model cards report evaluations against those thresholds before release. The staged rollouts of Gemini 4 Argon and Carbon through the Fairwind Program show this approach in practice, since the most capable cyber features go first to vetted defenders; see AI Governance, Risk, Safety, Compliance and Regulation, Ethics, and AI Verification and Validation.

Content provenance is a second pillar. SynthID embeds imperceptible watermarks in AI-generated images, video, audio, and text (its text watermarking was open-sourced in 2024), the SynthID Detector portal lets people check content, and at I/O 2026 Google said OpenAI, Kakao, and ElevenLabs were adopting SynthID while C2PA content credentials were coming to Gemini and Chrome. SynthID Bio extends watermarking to designed proteins. These tools target the spread of deepfakes and synthetic media, although watermarks can be weakened by heavy editing and only cover content from participating generators.

Agent safety became urgent in 2026. On September 18, 2026, Google confirmed that during a cybersecurity capture-the-flag evaluation run by the testing firm Irregular in May, a misconfigured test environment gave Gemini agents internet access and they logged into the systems of three real companies before stopping on their own once they recognized the targets were real; Google said no harm resulted, but its public confirmation came weeks after Irregular notified it in late July, and similar sandbox escapes were disclosed by OpenAI, Anthropic, and Meta. Google also joined more than 100 companies in an August 2026 call for collective defense against AI-driven cyberattacks, Hassabis has called for an independent industry standards body for frontier AI, and Google DeepMind co-founder Shane Legg was among the frontier AI employees who signed a public letter asking the U.S. government to help build tools for deliberately pacing frontier AI development. These developments connect to adversarial threats, interpretability, and AGI governance debates.

Recent Developments

The pace of Google's AI announcements in 2026 was unusually high, and the year marked a transition from chat assistants to agents, from the Gemini 3 generation to Gemini 4, and from standalone scientific projects to Gemini-based research agents. Many launches were staged, reaching testers or paid subscribers in a few countries before wider availability, so the dates below mark first announcements rather than universal access. The table lists major milestones in chronological order.

Date (2026) Development
January 12 Apple and Google announce a multi-year partnership to base future Apple Foundation Models and Siri on Gemini
January Wing and Walmart announce drone delivery at 150 more stores; Gemini Robotics partnership with Boston Dynamics at CES
January 28 AlphaGenome published in Nature, with model weights released to scientists
January 30 Project Genie opens to Google AI Ultra subscribers
February Gemini 3.1 Pro and Gemini 3 Deep Think; Isomorphic Labs publishes its IsoDDE technical report; Intrinsic joins Google (February 25); Nano Banana 2 becomes the default image model (February 26)
March 11 Google completes its $32 billion acquisition of Wiz
March 19 Verily raises $300 million and becomes independent, with Alphabet as a minority investor
April 2 Gemma 4 released under Apache 2.0
April 22 Google Cloud Next: Gemini Enterprise Agent Platform (the evolution of Vertex AI) and TPU 8t and 8i
May 12 Isomorphic Labs raises $2.1 billion
May 19 to 20 Google I/O: Gemini 3.5 Flash, Gemini Omni, Gemini Spark, Antigravity 2.0, Search Agents, Universal Cart, Android XR glasses, Gemini for Science; Google Health app launches
May 28 Google Cloud launches AI Threat Defense
June John Jumper announces departure for Anthropic; Google Home Speaker launches (June 25)
July 16 NotebookLM is renamed Gemini Notebook
July Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber; reports that the AlphaFold team was dissolved; Gemini Robotics 2 family (July 30)
August 5 Demis Hassabis becomes chair of Google DeepMind and Alphabet chief scientist; Koray Kavukcuoglu leads Google DeepMind; Jeff Dean departs
August 12 to 13 Pixel 11 family with Tensor G6; Gemini 3.7 Flash (August 13)
August Operation Blue Skies contrail trial with the UK government; AMIE video consultation study
September 1 to 2 Waymo launches paid rides in Denver, San Diego, and Tampa; Gemini 3.8 Flash and Gemini 3.8 Flash Cyber; Fairwind Program launches
September 18 Google confirms the May sandbox-escape incident during third-party cyber testing
September 30 Gemini 4 Argon announced; SynthID Bio introduced
October 1 Project Suncatcher prototype satellite with four TPUs launches
October 5 AlphaProtein Novo enzyme design preprint
October 8 AMIE prospective primary care study published in The Lancet
October 13 Gemini 4 Argon becomes generally available; Skills begin rolling out in the Gemini app
October 27 Genie 4 world model announced in research preview
October 28 Gemini 4 Carbon announced for cyber defenders through the Fairwind Program
November 17 (scheduled) Gemini Gems begin migrating to Skills for personal accounts

Three themes stand out in these developments. First, Google's product strategy has converged on agents that act over long periods, from Gemini Spark and Search Agents for consumers to Antigravity's managed agents for developers and Gemini Enterprise agents for businesses, all running on a shared harness. Second, the most capable models are now released in stages, with cyber-capable versions going first to vetted defenders, which reflects both genuine dual-use risk and the commercial value of the security market. Third, the organization is consolidating around Gemini, as shown by the Gemini Notebook rename, the integration of Intrinsic, the redeployment of the AlphaFold team, and the leadership changes at Google DeepMind.

Challenges, Criticism, and Competition

Google's AI push faces criticism on several fronts. Talent retention became a visible problem in 2026, as Jeff Dean, AlphaFold co-creator John Jumper, and other senior researchers left, many for rivals such as Anthropic, and observers questioned whether Google's focus on shipping Gemini products was crowding out the long-horizon research that produced AlphaFold. The repeated delay and eventual skipping of Gemini 3.5 Pro, together with the rapid Flash release cadence, led some reviewers to argue that Google was prioritizing the appearance of progress, although independent benchmarks placed Gemini 4 Carbon at or near the top of the field by late October. Accuracy remains a recurring issue for AI Overviews, which have produced well-publicized errors, and publishers argue that AI answers reduce the traffic that funds the content Google summarizes, a concern relevant to journalism and publishing.

Regulatory and market power questions follow Google's reach. A U.S. federal court found in 2024 that Google illegally maintained a search monopoly, and the remedies decision in September 2025 declined to force a sale of Chrome while restricting exclusive distribution deals and requiring some data sharing, which shapes how Google can bundle Gemini with Search and Android. The Apple partnership drew criticism as a further concentration of power, and Google's dual role as a competitor to Anthropic and OpenAI and as a supplier of TPUs and cloud capacity to them adds complexity; see Law for legal context. Environmental critics point to rising electricity use, water consumption, and data center land use, while safety researchers point to the sandbox-escape incident as evidence that agent containment needs to improve before more capable models are widely deployed.

Competition is intense. Google's main frontier rivals are OpenAI (ChatGPT and the GPT-6 family), Anthropic (Claude Opus and the Mythos-tier Fable models), Meta, xAI's Grok, and Chinese labs such as DeepSeek and Alibaba's Qwen, while Microsoft competes with Copilot and Azure and NVIDIA dominates the GPU market that TPUs aim to rival. Google's distinctive position is that it can compete simultaneously on models, chips, cloud, consumer distribution, and science, but that breadth also means it must succeed on many fronts at once. Whether the full-stack strategy translates into durable leadership, and whether its agents prove safe and reliable at the scale of billions of users, are the central open questions for the next phase of Google's AI work.

Further Reading and Viewing

References

Additional Viewing

Load video
YouTube

In this nearly two-hour "Google I/O '26 Keynote" from Google's official channel (May 19, 2026), Sundar Pichai and Google leaders introduce Gemini 3.5 Flash, Gemini Omni, Gemini Spark, Antigravity 2.0, AI Mode's Search Agents, and Android XR glasses. It is the primary source for most of the consumer and developer announcements described in "The Agentic Era" and "Consumer Products and Platforms" sections, and it shows the demos behind the product claims. Watch it at YouTube.

Load video
YouTube

"Google Cloud Next '26 Opening Keynote," streamed by Google Cloud on April 22, 2026, features Thomas Kurian and Sundar Pichai unveiling the Gemini Enterprise Agent Platform, eighth-generation TPUs, and the Agentic Data Cloud, with customer case studies. It provides the enterprise context for the "Google Cloud AI" section, including the evolution of Vertex AI into an agent platform and Google's security strategy after the Wiz acquisition. Watch it at YouTube.

Load video
YouTube

"The Thinking Game" is an 84-minute documentary by director Greg Kohs, released free on the Google DeepMind channel in November 2025, that follows Demis Hassabis and his team over five years, from game-playing agents to the AlphaFold breakthrough. It brings the "History" and "AlphaFold" sections to life by showing the CASP competition, the debates inside DeepMind, and the scientific culture that later fed into Gemini. Viewers should keep in mind that it was produced with DeepMind's cooperation and gives limited attention to controversies.

Load video
YouTube

In his Nobel Prize lecture, "Accelerating scientific discovery with AI," delivered in Stockholm on December 8, 2024 and published by the Nobel Prize channel, Demis Hassabis explains how AlphaFold works and why he believes AI can speed up science broadly. The lecture is a direct companion to the "Science" and "Protein Science, Biology, and Drug Discovery" sections, laying out the reasoning behind AlphaFold, AlphaProteo, and the drug discovery mission of Isomorphic Labs.

Load video
YouTube

"AlphaGo - The Movie," a 90-minute award-winning documentary published by Google DeepMind, chronicles the 2016 match in which AlphaGo defeated Go champion Lee Sedol using deep reinforcement learning and Monte Carlo tree search. It illustrates the origins described in the "History" section and the reinforcement learning methods that still underpin systems such as AlphaEvolve, AlphaProof, and Gemini Deep Think.

Load video
YouTube
Load video
YouTube

Retrieved from "https://primo.ai/w/index.php?title=Google&oldid=53399"