Google I/O 2026: Everything Google Announced — and the 93 Agents That Built an OS in 12 Hours

Halfway through the Google I/O 2026 keynote, a Google DeepMind engineer named Varun Mohan gestured at a screen and explained that the operating system running on it had not been written by a person. Antigravity 2.0, Google’s new agent platform, had built it from scratch — in 12 hours, by spinning up 93 separate sub-agents, processing billions of tokens, and finishing the job for under $1,000 in compute.

That single demo tells you everything about I/O 2026. This was not a conference about a smarter chatbot. It was a conference about software that does the work on its own. Almost every announcement Google made on May 19 — and there were a lot of them — points in exactly the same direction: the agent era, shipped to billions of users on day one. Here is the full rundown, and the strategy hiding underneath it.

Google didn’t lead with its smartest model

The headline model was Gemini 3.5 Flash, and the most revealing thing about I/O 2026 is what Google said about it. Sundar Pichai described 3.5 Flash as sitting at “almost 90%” of frontier performance — almost, not above. Google did not walk on stage and claim to have the smartest AI in the room. It claimed to have the cheapest, fastest one that is good enough.

And 3.5 Flash is genuinely fast. Google says it runs roughly 4× faster than rival frontier models measured in output tokens per second, while still beating the older Gemini 3.1 Pro on coding, agentic, and multimodal benchmarks. It began rolling out the same day across the Gemini app, Search, the Gemini API, and Antigravity — no staggered launch, no waitlist.

The actually-smartest model, Gemini 3.5 Pro, is still in testing and ships next month. The ordering is the point. In an agent world, where a single task fires off dozens of model calls, speed and price per call matter more than topping a leaderboard. Pichai even conceded the competitive landscape is “very dynamic” — a notable hedge in a year when OpenAI’s GPT-5.5 and Anthropic’s Mythos have reset what “leading” means.

Google also introduced Gemini Omni, a multimodal model that takes image, audio, video, and text as input and produces editable video as output, grounded in real-world knowledge. It’s headed to the Gemini app, Google Flow, and YouTube Shorts.

The scale behind the bet

Google can afford to compete on price and distribution because the distribution is staggering. The numbers Pichai put on screen:

• The Gemini app has gone from 400 million monthly users at last year’s I/O to over 900 million today — more than doubling in a year.
• AI Mode in Search passed 1 billion monthly users in a single year. AI Overviews is over 2.5 billion.
• Google now processes 3.2 quadrillion tokens a month — a 7× jump in a year, up from 9.7 trillion two years ago.
• 8.5 million developers build with Google’s models every month.

When you already reach billions of people, you don’t strictly need the smartest model. You need a good one you can switch on everywhere at once. That is precisely what Google did with 3.5 Flash, and it is the quiet thesis of the entire event.

Gemini Spark: the agent that works while you sleep

The consumer headline was Gemini Spark, which Google flatly calls “your personal agent.” Spark is a cloud-based agent that keeps running after you close your laptop or lock your phone. Google’s framing was deliberate: this transforms Gemini “from an assistant that can answer your questions into an active partner that does real work on your behalf and under your direction.”

The examples were concrete. Spark can parse your credit card statement to surface forgotten subscriptions, watch your school emails for upcoming deadlines, or turn scattered meeting notes into a polished Doc and a draft reply. It integrates with Gmail, Docs, and the rest of Workspace today, and expands to third-party tools through MCP over the summer. Spark enters beta next week for Google AI Ultra subscribers in the US.

Riding alongside it, Daily Brief sifts your Gmail, Calendar, and Tasks into a single prioritized digest of the day ahead, suggesting next steps. It starts rolling out today to AI Plus, Pro, and Ultra subscribers in the US.

Search stopped being a search box

Search got its biggest rethink in years. AI Mode now runs on Gemini 3.5 Flash. The box itself is now an intelligent search box that expands as you type, built for long, conversational queries instead of three-word keywords, with AI suggestions that anticipate intent rather than just autocomplete.

More radical are the information agents — agents that run in the background 24/7, scanning blogs, news, social posts, and Google’s freshest live data to monitor whatever you asked them to watch: a price, a sports situation, a developing story. They arrive this summer for AI Pro and Ultra. Search will also soon generate mini apps: custom dashboards and trackers it builds on the fly for ongoing tasks. And a new Universal Cart, built on Google Wallet, unifies shopping across Gemini, YouTube, and Gmail — flagging price drops, price history, stock alerts, and even product incompatibilities when you’re, say, building a PC from parts across retailers.

Antigravity 2.0 and the 12-hour operating system

For developers, Antigravity 2.0 is the centerpiece — a standalone desktop application built entirely around orchestrating agents, paired with an Antigravity CLI, an SDK, Managed Agents in the Gemini API, and enterprise support through the Gemini Enterprise Agent Platform. Subagents run inside cross-platform terminal sandboxes with credential masking and hardened Git policies.

This is the platform behind the 12-hour-OS demo. The pitch is blunt: complex software becomes a matter of dispatching and supervising a swarm of agents rather than writing code line by line. Whether that holds up outside a keynote is the open question of the year — but the demo was real, and it was the loudest moment of the show.

Two web standards landed too. WebMCP, a proposed open standard, lets any website expose structured tools that a browser’s AI agent can call directly — no scraping, no brittle automation — with an experimental origin trial starting in Chrome 149. Conceptually it looks like this:

// WebMCP (proposed) — a page hands structured tools to a browser agent
navigator.modelContext.registerTool({
  name: "book_appointment",
  description: "Book an available appointment slot",
  // the browser's agent can now call this directly
});

The new HTML-in-Canvas API lets developers drop real, interactive DOM elements into a WebGL/WebGPU canvas for immersive 3D interfaces.

Android 17 is now an “intelligence system”

Google reframed its mobile OS outright: Android 17 is being called an “intelligence system,” with Gemini running on-device and reaching across apps. The standout features:

• Create My Widget — describe the widget you want in plain language and Gemini builds it.
• App Bubbles — floating app windows for genuine desktop-style multitasking on foldables and tablets.
• Pause Point — mark distracting apps and Android shows a 10-second pause screen before it lets you in.
• Android Halo — a subtle strip at the top of any screen showing what your agent is doing in real time, so you can watch progress without stopping what you’re doing.

Switching from an iPhone now carries over far more, including saved passwords and your home screen layout. The Gemini app also picked up a redesign called Neural Expressive — fluid animations, haptics, new typography, and an inline Gemini Live that no longer hijacks the full screen.

Smart glasses, finally

After years of teasing, Android XR glasses got a real date. Google’s “intelligent eyewear” — audio-first glasses to start — ship this fall, with hardware from Samsung and Qualcomm and frames designed by Gentle Monster and Warby Parker. Notably, they pair with both Android phones and iPhones — a rare cross-platform olive branch from Google, and a sign it wants the form factor to win more than it wants to fence it off.

Workspace, YouTube, and the rest

The Gemini-everywhere push touched nearly every product Google ships. Gmail Live lets you conversationally search your inbox; Docs Live lets you create and edit documents by talking — both for AI Pro and Ultra this summer. Google Keep gets a mode that organizes free-flowing voice notes into clean, concise notes. Google Pics is an entirely new AI image generation and design app. And Ask YouTube handles complex questions and follow-ups across the whole catalog, returning a structured, interactive answer rather than a list of links (Premium subscribers, US, via youtube.com/new).

The price of agents

Agents burn compute, so Google rebuilt its pricing to match. Google AI Ultra dropped from $250 to $100 a month — with 5× the Gemini app usage of AI Pro — and the former $250 tier became a $200 tier with the same capabilities as before. The Gemini app is also moving off daily prompt caps to a “compute-used” model: a simple text prompt costs very little, a complex video or coding prompt costs more, and your limits refresh every five hours up to a weekly ceiling.

It’s an honest concession. In an agent world, “messages per day” is a meaningless unit. What actually matters is how much computation your agents consume on your behalf — so that’s what Google now meters.

When you can actually get it

Here is the quick-reference timeline for the major announcements:

+---------------------------+-------------------------------------+----------------------------+
| Announcement              | What it is                          | Availability               |
+===========================+=====================================+============================+
| Gemini 3.5 Flash          | Fast, cheap, near-frontier model    | Live now                   |
| Gemini 3.5 Pro            | The smartest model                  | Next month                 |
| Gemini Omni               | Multimodal video-creation model     | Rolling out (paid tiers)   |
| Gemini Spark              | 24/7 personal agent                 | Beta next week (Ultra, US) |
| Daily Brief               | Personalized morning digest         | Live now (US)              |
| Search information agents | 24/7 background monitoring          | This summer (Pro/Ultra)    |
| Universal Cart            | Cross-app agentic shopping          | This summer (US)           |
| Antigravity 2.0           | Agent-first developer platform      | Live now                   |
| Android 17                | Gemini-native "intelligence system" | Later this year            |
| Android XR glasses        | Audio-first intelligent eyewear     | This fall                  |
+---------------------------+-------------------------------------+----------------------------+

The verdict: distribution beats genius

Watch the whole keynote and a strategy emerges that is bolder than any single feature. Google is not betting it will own the smartest model. It is betting that a good-enough model, shipped instantly to 900 million Gemini users, a billion AI Mode users, and 8.5 million developers, beats a marginally smarter model that reaches far fewer people.

That is a real bet, and it carries a real weakness. Pichai’s “almost 90%” line is a tell. If an agent chains dozens of model calls to finish one task, small per-call quality gaps compound fast: a step that is 90% reliable is nowhere near 90% reliable across a 20-step workflow — it’s closer to one in eight finishing cleanly. Google is wagering that speed, price, and sheer distribution paper over that gap. Its competitors are wagering they don’t, and that reliability at the frontier is the only thing that matters once agents run unsupervised.

Demis Hassabis told the room that artificial general intelligence is “just a few years away.” Whether or not that lands on schedule, I/O 2026 made Google’s nearer-term thesis impossible to miss: the chatbot was the demo. The agent is the product. And on May 19, Google shipped it to everyone at once.

Found this useful? Follow me for daily LLM and AI deep-dives — I break down the news that actually matters for the people who build with these tools.