By the time Sundar Pichai walked off the Shoreline Amphitheatre stage on the evening of May 19, 2026, the word “assistant” had been quietly retired from Google’s vocabulary. In its place stood something far more ambitious: the agent.
For a decade, Google had promised that artificial intelligence would transform the way humanity works, searches, creates, and communicates. That promise had been made in developer keynotes, in quarterly earnings calls, in the quiet language of research papers published by Google DeepMind. But for most people, the ones who used Search to look something up, who relied on Gmail to manage their days, who watched YouTube to pass the time — AI had remained, at its core, a very sophisticated question-and-answer machine.
Google I/O 2026 changed that.
Held on May 19 and 20 at the company’s Mountain View campus, the two-day developer conference served as the formal declaration of what Google is now calling the “Agentic Gemini Era,” a pivot from AI that responds to prompts toward AI that acts on behalf of its users, continuously, autonomously, and across every surface of digital life. The announcements that poured out of the keynote were staggering in their scope: new models, a rebuilt search experience, a persistent AI agent that runs on cloud servers around the clock, a pair of smart glasses, an entirely new approach to shopping, and a restructuring of the company’s subscription business. More than one hundred discrete announcements were made across the conference’s two days.
The scale of what Google revealed was not merely technical. It was a statement of competitive intent, a signal to the rest of the AI industry that the company with two and a half billion Search users and thirteen products used by more than a billion people each intends to convert that distribution into dominance.
A Decade in the Making
Pichai opened the keynote by invoking history. Ten years had passed since Google stood on this same stage and declared itself an artificial intelligence company first and a search engine second. The industry had been skeptical then. It was not skeptical now.
The numbers Pichai delivered in his opening remarks were staggering. Google’s model APIs were processing approximately nineteen billion tokens per minute. More than 8.5 million developers were building applications with Google’s models each month. Over the previous twelve months, more than three hundred and seventy-five enterprise customers had each processed more than one trillion tokens, a unit of computational measurement so large that Pichai acknowledged the absurdity of the moment. “Never imagined I’d say quadrillion in an I/O keynote,” he said, “but here we are.”
AI Mode in Google Search had crossed one billion monthly users. AI Overviews, the feature that generates summarized answers at the top of search results, were now reaching two and a half billion users every month across more than two hundred countries and ninety-eight languages.
These were not projections. They were reported figures from a company that, for all of the competitive noise surrounding OpenAI, Anthropic, Meta, and Microsoft, still operates the most widely used set of digital tools on the planet. Pichai’s subtext was clear: no competitor could match the distribution advantage Google held, and Google intended to use every inch of it.
The New Engine: Gemini 3.5 Flash
At the center of every announcement made at I/O 2026 was a single model: Gemini 3.5 Flash.
Released the same day as the keynote and made immediately available to consumers through the Gemini app, to developers through the Gemini API and Google AI Studio, and to enterprises through Gemini Enterprise Agent Platform, the model represents Google’s most significant model launch in recent memory. Not because of its raw intelligence on any single benchmark, but because of the combination of speed, capability, and cost it delivers simultaneously.
Google described 3.5 Flash as the first model in a new family designed to deliver what the company calls “frontier intelligence with action” — a phrase that signals a deliberate departure from the previous generation of models, which were evaluated primarily on their ability to answer questions well. The 3.5 series is built to do things, not just answer things.
On benchmark performance, the numbers are notable. Gemini 3.5 Flash scores 76.2 percent on Terminal-Bench 2.1, which evaluates coding performance for AI agents. It achieves 1,656 Elo on GDPval-AA, which measures real-world agentic task performance. It scores 83.6 percent on MCP Atlas, which tests multi-step tool-use reliability, and 84.2 percent on CharXiv Reasoning, which evaluates multimodal understanding. On all four of those benchmarks, 3.5 Flash outperforms Gemini 3.1 Pro, the previous generation premium model. A smaller, faster, cheaper model now outperforms last year’s flagship.
On output speed, Google claims 3.5 Flash runs four times faster than comparable frontier models, measured in output tokens per second. The model carries a context window of one million tokens, accepts text, images, audio, video, and PDF input, and produces up to 64,000 output tokens. API pricing is set at $1.50 per million input tokens and $9.00 per million output tokens. It is free for users of the Gemini app and AI Mode in Google Search.
It is worth noting that the benchmarks cited are Google’s own self-reported evaluations. Independent third-party validation had not been completed at the time of publication. A separate evaluation from Appwrite Arena, which tests across 191 questions in enterprise service categories, found that 3.5 Flash still trails Gemini 3.1 Pro on certain tests, including Humanity’s Last Exam and ARC-AGI-2. The model is exceptional at what it was designed for, which is agentic and coding work, but it is not a clean replacement for the Pro tier across every dimension.
Gemini 3.5 Pro, the more powerful sibling, was confirmed as already in use internally at Google and is expected to be released to the public in June 2026.
Gemini Omni: Teaching Machines to Understand the World
If 3.5 Flash was the workhorse of I/O 2026, Gemini Omni was its showpiece.
Introduced on stage by Demis Hassabis, the CEO of Google DeepMind, Gemini Omni is a fundamentally different kind of model. Rather than being built for speed or cost efficiency, Omni is built for what Google calls world understanding: the ability to accept virtually any kind of input and generate virtually any kind of output, with a particular emphasis on video creation grounded in the laws of physics.
Hassabis demonstrated the model’s capabilities by prompting it to generate a clay animation video explaining the process of protein folding, a topic of considerable complexity. The model produced a video that accurately depicted molecular behavior while framing the science in an accessible, visual narrative. It was the kind of output that, even twelve months ago, would have required a production team, a subject matter expert, and hours of animation work.
The technical underpinnings of Gemini Omni reflect the ambitions of the model. It combines the reasoning capabilities of the Gemini family with Google’s generative media models, including Veo for video and Nano Banana for image generation. The result is a model with what Google describes as an improved understanding of physical forces, including gravity, kinetic energy, and fluid dynamics. Omni-generated videos are grounded in what Google calls “real-world knowledge,” meaning the model does not simply generate visually plausible content but attempts to generate content that is physically accurate and semantically coherent.
The first version released at I/O, Gemini Omni Flash, accepts images, audio, video, and text as input and outputs video. Google’s stated ambition for the Omni family is broader: over time, the company intends for it to generate any output from any input. Every video created by Omni carries SynthID, Google’s imperceptible digital watermark, which can be verified through the Gemini app and a new dedicated portal called SynthID Detector.
Gemini Omni is available to paid subscribers in the Gemini app, in Google Flow for creative professionals, and in YouTube’s Shorts Remix and Create features for users aged eighteen and older.
Hassabis, speaking at I/O and in a subsequent interview with Fast Company, repeated his prediction that artificial general intelligence was approximately four years away. He described AGI’s potential impact as ten times that of the Industrial Revolution, occurring at ten times the speed. The claim is extraordinary and contested — Andrew Ng, one of the founders of Google Brain, believes the arrival of AGI is decades away — but it reflects the conviction driving the research agenda at DeepMind and shapes the ambition behind every product announced at I/O 2026.
Reimagining Search: The Biggest Upgrade in Twenty-Five Years
Google Search has been the foundation of Google’s business since the company was founded in 1998. For most of that time, its fundamental interface remained unchanged: a rectangular text box, a blinking cursor, and a list of ten blue links in response.
At I/O 2026, Google announced what it called the biggest upgrade to the Search box in over twenty-five years.
The new Search interface is built around a fundamentally different premise. Rather than asking users to compress their questions into keywords, it invites them to express full thoughts, complex intentions, and multimodal inputs. The Search box now dynamically expands as users type, accommodating longer, more conversational queries. Users can submit not just text but images, videos, files, and even open browser tabs as inputs. AI-powered query suggestions go beyond autocomplete, anticipating intent rather than simply completing phrases.
AI Mode, the AI-powered layer of Search that Google has been building out over the past year, is now powered by Gemini 3.5 Flash and serves as the default experience in over two hundred countries. The feature allows Search to engage in multi-turn conversations, build on previous queries, and synthesize answers from across the web rather than simply ranking and returning links.
Beyond conversational improvements, Google announced a feature called Generative UI, which allows Search to construct entirely custom response formats on the fly. Using Antigravity, Google’s developer platform, Search can assemble interactive visuals, tables, graphs, and simulations in real time, designed specifically around the shape of the user’s question. A query about the orbit of a comet might generate an animated simulation. A query about historical inflation might produce an interactive chart. Google confirmed this feature will be available to all users, free of charge, beginning in the summer of 2026.
Perhaps the most consequential Search announcement was the introduction of Information Agents, a new category of AI capability designed to monitor the web on a user’s behalf. Rather than requiring users to return to Search to check on ongoing topics, Information Agents run continuously in the background, scanning news sites, blogs, social media, financial data, sports scores, and shopping trends, and sending users synthesized updates when something relevant changes. A user planning to buy a particular product can instruct an agent to monitor for a price drop. A person tracking developments in a legal case can instruct an agent to surface new filings. The agents are described as drawing from the full breadth of Google’s real-time data infrastructure.
Information Agents will be available to Google AI Pro and Ultra subscribers beginning in the summer of 2026.
Gemini Spark: The Agent That Never Sleeps
The single announcement at I/O 2026 that generated the most sustained discussion among technology observers was Gemini Spark.
Spark is a personal AI agent designed to run continuously on Google Cloud virtual machines, twenty-four hours a day, seven days a week, regardless of whether the user’s devices are on or off. Unlike a chatbot, which waits for a prompt before doing anything, Spark is designed to operate on its own initiative within the parameters a user defines, taking actions, tracking tasks, communicating with services, and reporting back.
During the keynote, Google demonstrated Spark’s capabilities through a scenario involving a neighborhood block party. Given a set of unstructured planning materials, Spark pulled together RSVPs from Gmail, tracked contribution commitments across a spreadsheet in Google Sheets, drafted follow-up emails to neighbors who had not responded, and assembled a presentation in Google Slides complete with logistical details sourced from a document in Google Drive. The demonstration was notable not for any single action Spark took but for the fact that it took all of them in sequence, without being re-prompted at each step.
Spark is built on Gemini 3.5 Flash and runs through Antigravity, Google’s agent development platform. It integrates natively with Gmail, Google Docs, Sheets, Slides, Calendar, and other Workspace applications. Google has confirmed that Model Context Protocol support for third-party applications, including tools like Canva, OpenTable, and Instacart, is planned for later in the summer of 2026. Accessible through the Gemini app, Spark can also be reached by email or direct message, an unconventional design choice that reflects the assumption that not every interaction with an AI agent needs to happen inside an app interface.
Google has confirmed that users must manually approve any sensitive actions, including sending emails, making purchases, or modifying calendar events. Spark will not act without authorization on tasks that carry meaningful real-world consequences.
The Gemini app itself received a comprehensive redesign to accompany these new capabilities. Google introduced what it calls a “Neural Expressive” design language, featuring fluid animations, vibrant colors, haptic feedback on mobile, and updated typography. Support for regional dialects in Gemini’s voice responses, described by Google as covering a wide range of linguistic variations globally, is expected to follow in the coming months.
Gemini Spark is available to Google AI Ultra subscribers in the United States and will expand internationally as capacity allows.
Daily Brief and Workspace: Bringing AI Into the Work Day
Alongside Spark, Google announced a related but lighter-weight feature called Daily Brief, described as a personalized AI-generated summary of each user’s day.
Drawing from Gmail, Google Calendar, and Tasks, Daily Brief synthesizes a morning overview that organizes what the user needs to know, prioritizes action items, and suggests next steps. It is, in essence, the first thing Spark might say to a user each morning if Spark were summarizing overnight activity. Daily Brief launched at the time of the keynote for Google AI Plus, Pro, and Ultra subscribers in the United States.
For Google Workspace users, the most significant announcement was a feature called Docs Live. Designed to eliminate the friction between thinking and writing, Docs Live allows users to dictate documents in natural, conversational speech, including false starts, mid-sentence corrections, and colloquial phrasing, and generates a polished, formatted Google Doc from the audio. The system handles the translation from spoken to written language, applying structure and formatting automatically. Docs Live is expected to roll out to AI Pro and Ultra subscribers in the summer of 2026, with similar capabilities coming to Gmail and Google Keep.
Gmail is also gaining a feature called Gmail Live, which allows subscribers to have natural language conversations with their inbox. Rather than searching through threads manually, users can ask questions like “Has anyone responded about the budget proposal?” or “What did the client say last month?” and receive synthesized answers. Gmail Live is planned for AI Pro and Ultra subscribers beginning in the summer.
Search Meets Commerce: The Universal Cart
One of the more strategically revealing announcements at I/O 2026 was not about models or agents. It was about shopping.
Universal Cart is Google’s attempt to build a unified, intelligent commerce layer across its entire product ecosystem. Rather than requiring users to visit individual merchant websites, add items to separate carts, and manage multiple checkout flows, Universal Cart allows users to accumulate items across their Google experience: while browsing Search, while chatting with Gemini, while watching YouTube, or while reading an email in Gmail. A single cart aggregates all of it.
The intelligence built into Universal Cart goes beyond simple aggregation. Once an item is added, the system monitors it continuously. It searches for deals and price drops, tracks price history, compares across merchants, and alerts the user when a product that was out of stock becomes available. The cart operates in the background, effectively functioning as a passive shopping assistant that never forgets what the user was considering.
Universal Cart will begin rolling out across Search and the Gemini app in the United States during the summer of 2026, with YouTube and Gmail integration to follow.
Ask YouTube and Google Pics: Two New Surfaces
Two additional product announcements broadened the scope of Google’s AI ambitions.
Ask YouTube, now available for YouTube Premium subscribers in the United States through youtube.com/new, transforms the search and discovery experience on the world’s largest video platform. Rather than requiring users to browse lists of results, Ask YouTube allows conversational queries and follow-ups, surfacing the most relevant videos from across YouTube’s catalog with structured, interactive responses. The feature represents a significant departure from keyword-based video search and reflects Google’s broader strategy of making every surface in its ecosystem conversationally accessible.
Google Pics is a standalone image creation and editing application built on the company’s Nano Banana model. Designed for users who want to create visual content without technical expertise in image generation, Pics supports everything from designing promotional materials to generating custom graphics, with editing controls exposed through a simplified interface. The application is aimed at the broad middle of the market, users who want capable image generation without the learning curve associated with more complex tools.
Android XR and the Return of Smart Glasses
In one of the more visually dramatic moments of the I/O keynote, Google revealed the first tangible hardware to emerge from its Android XR platform: a line of smart glasses designed for everyday wear.
The glasses come in two configurations. The first is an audio-focused form factor featuring cameras, microphones, and speakers, designed for all-day wear. The second adds an optional in-lens display for users who want visual overlays alongside audio capabilities. Both configurations are built on hardware developed in partnership with Samsung and Qualcomm, with exterior designs by Gentle Monster and Warby Parker. Google confirmed that the glasses are compatible with both Android and iOS devices, a deliberate choice to maximize the addressable market.
The first Android XR smart glasses are expected to ship in fall 2026. Pricing had not been confirmed at the time of publication.
The glasses represent Google’s most serious entry into wearable AI hardware since the discontinuation of Google Glass, which struggled with both privacy concerns and a lack of practical utility. The company’s approach this time is more measured: functionality is anchored to practical tasks like getting directions, sending messages, and taking photos, rather than presenting a comprehensive augmented reality overlay that early Glass critics found overwhelming and socially awkward. Gemini is integrated throughout, providing a voice-accessible AI interface without requiring the user to reach for a phone.
Antigravity 2.0 and the Developer Ecosystem
For the developer community, the most consequential announcement at I/O 2026 may have been the launch of Antigravity 2.0.
Antigravity is Google’s agent-first development platform, the infrastructure layer that allows developers to build, orchestrate, and deploy AI agents at scale. Version 2.0 ships with a new command-line interface, new capabilities for multi-agent coordination, and deep integration with Gemini 3.5 Flash. The platform is now globally available at no cost.
Alongside Antigravity 2.0, Google launched Managed Agents in the Gemini API, powered by the Antigravity agent framework. The feature allows enterprise developers to create agents that can be orchestrated, monitored, and scaled through a standardized interface, lowering the complexity involved in building reliable agentic systems.
The developer keynote, held on the second day of the conference, framed these tools within a broader narrative: Google is no longer simply building AI models and exposing them through an API. It is building an end-to-end agentic development stack, from model through orchestration through deployment, designed to make any developer capable of building sophisticated multi-step AI workflows. As Pichai noted in his keynote, more than 8.5 million developers are already building with Google’s models monthly. Antigravity 2.0 is designed to give all of them a path from prototype to production agent.
Silicon and Infrastructure: The TPU 8 Generation
No discussion of Google’s AI ambitions is complete without addressing the infrastructure that makes them possible. At I/O 2026, Pichai announced the eighth generation of Google’s Tensor Processing Unit chips, the custom silicon that has powered Google’s AI training and inference workloads since the technology was first introduced a decade ago.
The eighth generation takes a dual-chip approach for the first time, splitting the hardware into specialized designs for different workloads. The TPU 8t is optimized for large-scale pretraining and delivers nearly three times the raw computing power of the seventh-generation chip. The TPU 8i is designed for inference, and is described as both faster and more energy efficient than its predecessor. Together, the two chips allow Google to distribute training workloads across data centers globally, removing the constraint that previously required large training runs to be confined to a single facility.
Separately, Google and Blackstone announced a significant new infrastructure partnership ahead of the conference, backed by a five-billion-dollar investment, aimed at expanding access to Google’s TPU chips and AI computing capacity to meet growing global demand.
The Price of Intelligence: A New Subscription Architecture
One of the more practically significant announcements at I/O 2026 had nothing to do with technology. It had to do with pricing.
Google restructured its AI subscription tiers, introducing a new one-hundred-dollar-per-month Google AI Ultra plan positioned between the existing twenty-dollar AI Pro plan and the top-tier plan, which was simultaneously reduced from $250 per month to $200 per month.
The new $100 tier includes five times the usage limits of AI Pro, access to Gemini Spark in beta, twenty terabytes of Google One storage, and the Google Cloud credits required to run agentic workloads through Antigravity. The $200 tier retains twenty times the Pro usage limits, thirty terabytes of storage, access to Project Genie for interactive world creation, and additional premium features including Project Mariner.
The pricing move is an implicit acknowledgment of competitive pressure. ChatGPT Pro from OpenAI had been priced at $200 per month, and Claude Max from Anthropic had entered the market at $100 per month. Google’s introduction of a $100 entry point to its highest-capability tier signals a willingness to compete on price at the premium end of the AI subscription market, while the reduction of the $250 plan to $200 narrows the gap with OpenAI further.
The restructuring was not without its complications. The AI Pro plan, which previously used a fixed-message system, was simultaneously moved to a credit-based system tied to computational complexity. A number of users on community forums reported that the new system felt more restrictive in practice, with single complex prompts consuming significantly more of their monthly allowance than expected. Google confirmed the change and stated that the credit-based system more accurately reflects the actual computational cost of complex queries.
The Competitive Horizon
I/O 2026 lands at a moment of extraordinary competition in the AI industry. OpenAI has released GPT-5.5 and continues to build out the ChatGPT platform. Anthropic has introduced Claude Opus and expanded Claude Code. Meta is distributing its Llama models broadly. Microsoft is embedding Copilot across its enterprise software. Apple is expected to make significant AI announcements at WWDC 2026 on June 8, with speculation running high about whether the company will integrate a third-party AI, possibly including Gemini, as the intelligence layer behind Siri.
What distinguishes Google’s position, as laid out across the two days of I/O 2026, is not any single model or any single feature. It is the combination of infrastructure, distribution, and product surface area that no competitor can replicate from scratch. Gemini is not competing for users in a chatbot marketplace. It is being embedded, quietly and pervasively, into the tools that billions of people already use every day.
The question the industry is now asking is not whether Google has the technology to compete at the frontier of AI development. The events of May 2026 settled that question convincingly. The question is whether the users on the other end of all that infrastructure — the ones who open Gmail in the morning, who search for a product recommendation, who ask YouTube to recommend a video — will notice the change, and whether, once they do, they will find their lives meaningfully different because of it.
If the scale of what Google announced at I/O 2026 is any indication, they are not waiting to find out.