Google I/O 2026

What You Will Learn

• What Google actually announced at I/O 2026 and which features were still in preview.
• How Gemini 3.5 Flash and Gemini Omni differ in purpose and output.
• What Gemini Spark, Daily Brief, and Search agents can do under user control.
• Why the old claims about Project Aura and an agent-only future need careful qualification.

What Google Announced at I/O 2026

Google I/O 2026 was built around what Google called the agentic Gemini era. The company’s CEO Sundar Pichai keynote transcript and official collection put three themes together. Google introduced Gemini 3.5 as a model family designed for intelligence and action, Gemini Omni as a model for multimodal creation that started with video, and a wider set of agents across Search, the Gemini app, shopping, productivity, and development.

The event was not one product launch. It was a portfolio presentation. Some products were generally available, some were rolling out to subscribers, and some were limited to trusted testers or announced for later in the year. That distinction is essential because the old article treated demos, previews, and availability as if they were all the same thing.

Google’s official I/O collection lists Gemini Omni, Gemini 3.5, Gemini Spark, Daily Brief, Information agents in Search, Universal Cart, Google Antigravity, and intelligent eyewear. The detailed announcement list adds Google Pics, Google Flow Agent, Ask YouTube, Docs Live, Gemini for Science, and other product updates.

The event is best read as a product roadmap with selected live launches. Readers looking for the larger history of AI product releases can compare it with the site’s multimodal AI analysis, but the official Google pages remain the source for I/O-specific claims.

Gemini App Passed 900 Million Monthly Users

Google said the Gemini app had passed 900 million monthly active users by I/O 2026. Pichai’s keynote transcript compared that figure with 400 million monthly active users at the previous I/O and said daily requests had grown more than seven times in the same period. Google’s Gemini app announcement described the audience as more than 900 million people across 230 countries and more than 70 languages.

This is a Google-reported usage figure. It is not an audited independent estimate in the retrieved sources, and it should not be confused with every person who has ever opened the app. The same keynote gave different numbers for other products. Google said AI Overviews had more than 2.5 billion monthly active users and AI Mode had surpassed 1 billion monthly active users.

The figures show Google’s distribution advantage, but they do not prove that users have adopted every new agent feature. Gemini Spark was still rolling to trusted testers, while some connected actions depended on subscription tier, geography, and user opt-in. Scale and feature availability are separate questions.

For a practical comparison of model families rather than product reach, see the site’s AI model guide. User counts tell readers about distribution. They do not by themselves establish model quality or safety.

Gemini 3.5 Flash Adds Action to the Model Family

Google introduced Gemini 3.5 Flash as the first model in its new 3.5 series. The company described it as combining intelligence with action and positioned it for coding, long-horizon tasks, and workflows that require a model to plan and use tools. Google’s official announcement list gives the company’s launch and benchmark claims. Google said the model was available through Google Antigravity, the Gemini API, Google AI Studio, and Android Studio.

Google also published benchmark results and speed comparisons. Its announcement list said Gemini 3.5 Flash outperformed Gemini 3.1 Pro on selected coding and agent benchmarks including Terminal-Bench 2.1, GDPval-AA, and MCP Atlas. Pichai’s transcript said Flash was four times faster than other frontier models on output tokens per second. These are vendor-reported comparisons, so the model, test conditions, and score definitions matter.

Flash is not the same thing as a general-purpose agent. The model supplies reasoning and tool-use capability, while a product such as Antigravity supplies an execution harness, permissions, files, and tools. A reliable workflow depends on all of those layers. A faster model can reduce waiting time without removing the need for review and access control.

Gemini Omni Focuses on Multimodal Video Creation

Gemini Omni was Google’s model for turning different input types into generated media. Google said Omni could use combinations of text, images, video, and initially voice references to create video outputs. The company said the model was starting with video and would expand to other output modalities over time.

Google positioned Omni as both a generation and editing system. The Gemini app announcement described conversational changes such as cinematic zooms, background replacement, and the use of footage from a camera roll. Google also said Omni Flash was rolling out to Google AI Plus, Pro, and Ultra subscribers through the Gemini app and Google Flow, with additional availability in YouTube Shorts and APIs to follow.

The practical point is that input and output are not the same as general understanding. A system can accept multiple media types while still having limitations in identity consistency, timing, physical detail, copyright handling, and safety. Google said generated Omni videos include a SynthID digital watermark, which can help verification but does not answer every question about provenance.

Gemini Spark Is a Controlled Personal Agent

Gemini Spark was presented as a 24/7 personal AI agent that can work on tasks under a user’s direction. Google said Spark runs on Gemini 3.5 and uses the Antigravity harness. The company described it as a cloud-based system that can keep working after a laptop is closed or a phone is locked, subject to the feature’s rollout and connected-app permissions.

Google’s Gemini app page gave examples such as recurring tasks, inbox monitoring, meeting-note synthesis, and drafting documents or emails. It also said Spark would connect with Workspace tools and later with third-party services through MCP. The important limitation is that Spark was not announced as a fully open consumer service. Google said it was rolling first to trusted testers and planned a Beta for U.S. Google AI Ultra subscribers.

Google also said Spark operates under user direction and is designed to ask before high-stakes actions such as spending money or sending emails. That is a product safeguard, not a guarantee that every workflow will be error-free. Users still need to choose connected apps carefully, review permissions, and check outputs before consequential actions.

The site’s Oxford lecture analysis discusses why action-taking systems raise different questions from chat interfaces. Spark is a useful example because it combines a model, connected services, and background execution rather than only generating text.

Daily Brief and the New Gemini Interface

Daily Brief was introduced as an out-of-the-box agent for the Gemini app. Google said it could gather information from connected apps, organize a user’s day, and suggest next steps. The Gemini app page said Daily Brief was rolling out to Google AI Plus, Pro, and Ultra subscribers starting in the United States.

Google also announced a design language called Neural Expressive. The company described fluid animations, new typography, haptic feedback, interactive visuals, timelines, and dynamic graphics. The Gemini app page said the redesigned interface was rolling out globally across web, Android, and iOS.

A more visual interface can make an AI answer easier to scan, but it can also make uncertain content look settled. Users should distinguish a polished display from a verified source. The official product page says availability varies by subscription tier and geography, so the experience will not be identical for every account.

AI Mode and Generative UI in Search

Google said AI Mode had surpassed 1 billion monthly active users and that Gemini 3.5 Flash would become its default model globally. It also announced Information agents that could monitor topics in the background and send synthesized updates with links. Google said those agents would begin rolling out in summer, starting with Google AI Pro and Ultra subscribers.

Another announcement was generative UI in Search. Google said Antigravity and Gemini 3.5 Flash could build custom layouts, tables, graphs, simulations, dashboards, and other interactive formats for a question. The company described some experiences as mini apps for ongoing tasks, with rollout timing dependent on the feature.

Generative UI changes how search results are presented, but it does not remove the need to inspect sources. A custom dashboard can organize information without making the underlying data current or correct. Readers should open the links, check dates, and look for the original authority behind a claim.

For broader context on how AI systems combine text, images, audio, and action, read the site’s multimodal systems coverage. Search agents are a product feature, not proof that every web answer has been verified.

Google Antigravity 2.0 and Managed Agents

Google Antigravity was described as an agent-first development platform. At I/O, Google announced Antigravity 2.0 as a standalone desktop application for orchestrating agents, along with a command-line interface and SDK. The announcement list also described Managed Agents in the Gemini API, which can provision a remote Linux environment where an agent can call tools, execute code, manage files, and browse for data.

These features are aimed at builders rather than ordinary chat users. The benefit is a structured execution environment that can hold files, tools, and task state. The risk is that the same environment can receive permissions that are difficult to review. Developers need sandbox boundaries, audit logs, secret controls, and explicit limits on network access.

The site’s AI research workflow article covers a related question. An agent that can perform work needs a review path for both its reasoning and its actions.

Universal Cart and Proactive Shopping

Google introduced Universal Cart as a shopping hub that can collect products from Search, Gemini, YouTube, and Gmail. Google said the cart could find price drops, show price history, alert users when an item returns to stock, and identify product incompatibilities. The system was described as using Gemini models and Google Wallet information such as payment benefits and loyalty details.

Google also announced the Universal Commerce Protocol for checkout flows. The company said Universal Cart would roll out across Search and the Gemini app in summer, with YouTube and Gmail to follow. Availability, merchants, payment methods, and geographic coverage may differ.

This is an example of an agent crossing from information into action. Product recommendations, account connections, and checkout all create a need for clear user control. A lower price does not necessarily mean the right product, and an automated cart should not be treated as a substitute for checking seller terms, delivery conditions, or return rules.

Google Pics, Flow Agent, and Creative Tools

Google announced Google Pics, an image creation and editing tool built on its Nano Banana model. The company described object-level editing, text editing, translation, and Workspace integrations. Google said Pics was initially limited to trusted testers and would roll out later to Google AI Pro and Ultra subscribers and Workspace customers.

Google Flow also received an agent that could help with brainstorming, creation, editing, asset organization, and batch changes. Google said Flow Agent was available to Flow users globally and that subscribers could create and remix custom tools. These are creative workflows with different access and rollout rules from Gemini Spark.

Google also announced SynthID verification and Content Credentials features across Gemini, Search, and Chrome. Google said SynthID had watermarked more than one hundred billion images and videos and sixty thousand years of audio assets, while its announcement list said verification had been used 50 million times globally. Those figures come from Google and should be attributed.

Android XR and Intelligent Eyewear

Google’s I/O announcements included intelligent eyewear for Android XR. The company described two types. Audio glasses would provide spoken assistance, while display glasses would show information in the user’s field of view. Google said its first audio glasses, made with partners including Gentle Monster, Warby Parker, and Samsung, would arrive in the fall and work with Android and iOS devices.

The official sources retrieved for this article did not establish the old article’s `Project Aura` name as the official I/O 2026 product label. The reliable wording is Android XR intelligent eyewear with audio and display glasses. Product names, partner plans, and release dates can change before hardware reaches buyers.

Eyewear also introduces privacy questions that do not apply in the same way to a phone screen. A camera, microphone, display, and assistant can operate around people who did not choose to use the device. Google’s announcement focused on hands-free assistance, but users and regulators will still need clear rules for recording, notification, consent, and data retention.

What Google I/O 2026 Means for Users and Developers

Google I/O 2026 was a shift in product design toward systems that can monitor context, plan across steps, and act through connected services. The company’s official claims are significant. The Gemini app passed 900 million monthly active users, AI Mode passed 1 billion monthly active users, and Gemini 3.5 Flash and Gemini Omni expanded the model portfolio.

The next question is not whether an agent can complete a demonstration. It is whether users can understand what it can see, what it can change, when it will ask permission, and how to stop it. Google’s product pages repeatedly describe staged rollout, trusted testers, subscription limits, opt-in connections, and confirmation before high-stakes actions.

The corrected view is more useful than the old claim of an immediate agent-only future. Google announced a set of models, agents, developer tools, and hardware plans. Some were available at I/O. Others were previews or scheduled for later. The practical work now is to test the features under clear permissions and measure whether the extra automation improves the user’s task without hiding important choices.