Google I/O 2026: Complete recap — Gemini 3.5, Spark, XR eyewear, and everything else

Context: where Google stands in the AI race today

Sundar Pichai opened the keynote with the framing “the last year was extraordinary” and showed numbers meant to underline the size of the operation. According to him, Google AI APIs process approximately 19 billion tokens per minute. Over the last 12 months more than 375 customers each processed over 1 trillion tokens. And 50 billion imageshave already been generated by the Nano Banana models — which Pichai described as last year's “breakout star.”

Never imagined I'd say quadrillion in an I/O keynote, but here we are.

Pichai repeatedly framed the whole of I/O 2026 as “the phase where people want to see value in the products they use every day.” That signalled a shift away from abstract benchmark numbers toward integration of AI into everyday workflow — confirmed by the demos that followed, which focused on Gmail, Docs, Search, and mobile devices.

Ten years since we pivoted the company to be AI-first, we still see AI as the most profound way to advance our mission and improve people's lives at scale.

Demis Hassabis, CEO of Google DeepMind, later added a line that flew around every tech magazine within an hour: “AGI is now on the horizon.” He underscored the importance of safety in agentic systems — which the keynote then linked directly to the detailed security features in Antigravity 2.0.

Gemini 3.5 Flash and 3.5 Pro

The headline release of I/O 2026 was Gemini 3.5 Flash. According to Google it combines “frontier intelligence with an agentic ability to complete tasks” and beats Gemini 3.1 Pro on coding, agentic, and multimodal benchmarks. Google claims 4× faster output token generation compared to other frontier models.

The model is rolling out immediately in the Gemini app, in AI Mode in Google Search, in Antigravity 2.0, and through the Gemini API. Specific API pricing was not announced on stage; third-party sources cite roughly $1.50 per million input tokens and $9.00 per million output tokens in the global region, but the exact numbers should be verified in the Google AI for Developers documentation.

Gemini 3.5 Prois in testing with a “next month” ship date. Details on parameters, pricing, and availability Google left for a later announcement.

Gemini Spark — personal agent

Spark is the new arrival that received the most stage time in the keynote. Google framed it as a shift from assistant to active partner: “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.”

Spark integrates deeply with Gmail, Docs, and other Google Workspace apps. Third-party tooling will be added in summer through the Model Context Protocol (MCP). Launch: next week for Google AI Ultra subscribers, US first.

On-stage live demo:the presenter used Spark to plan a wedding. In real time the agent tracked RSVP responses from Gmail, generated reminders for guests who had not replied, updated the guest list in Sheets, and proposed next steps. The point of the demo was to show an “autonomous background workflow,” not a one-shot answer.

Gemini Omni — multimodal model with editable video output

Omni is Google's answer to OpenAI Sora and Anthropic Claude with video capabilities. It combines reasoning with creation; accepts image / audio / video / text inputs. Key feature: the output is editable video “grounded in real-world knowledge.”

Live demo:the presenter took a video of a woman playing guitar and Omni automatically added cinematic visual effects — lighting, particle effects, depth of field. Editing was done in natural language (“highlight the guitar,” “add a soft golden glow”).

Availability: rolling out for AI Plus, Pro, and Ultra subscribers in the Gemini app, Google Flow, and YouTube Shorts. Companion apps Google Flow and Flow Music— Flow on Android in beta (iOS soon), Flow Music on iOS available (Android soon).

Gemini app redesign — Neural Expressive

The whole Gemini app gets a new design language called Neural Expressive: fluid animations, more vibrant colours, haptics, and new typography. The prompt box is now pill-shaped with a single “+” menu that integrates Tools.

The classic navigation drawer was replaced with a fullscreen interface. Gemini Live now has an inline experience — no need to switch to fullscreen. Responses surface the most important information at the top in bold, plus inline images, narrated videos, timelines, and interactive visualisations. Rolling out now to Android, iOS, and web.

Daily Brief:a new feature that generates a personalised daily digest — sifting through Gmail, Calendar, and Tasks, prioritising tasks, and suggesting next steps. Available today for AI Plus, Pro, and Ultra subscribers in the US.

Google AI subscriptions reorganization

Google substantially reworked subscription pricing. The main changes:

• AI Ultra has a new starting tier $100/month with 5× higher usage limits than AI Pro.
• The previous $250 plan is now $200, capabilities identical to before.
• Compute-Used Model: Google is dropping daily prompt limits. The new system meters prompt complexity, features used, and chat length. Limits refresh every 5 hours up to a weekly ceiling. Pichai justified it by saying that “a simple text prompt uses far less compute than a complex video or coding prompt.”

Google Search — AI Mode, Information Agents, and Universal Cart

AI Mode in Search now runs on Gemini 3.5 Flash (free). The new Intelligent Search Box expands as you type to reflect conversational queries, anticipates intent, and offers AI-powered query suggestions beyond classic autocomplete.

Information Agentsare a breakthrough for Search: agents that monitor the web 24/7 — blogs, news, social, real-time financial, shopping, and sports data — and alert you when something changes. Launch: summer 2026 for AI Pro and Ultra subscribers.

Mini Apps / Custom Dashboards:users can build their own dashboards and trackers for continuous tasks. “Upcoming months” for AI Pro and Ultra in the US.

Universal Cart: a Gemini-powered shopping agent available through the Gemini app, YouTube, and Gmail. Features: deal hunting, price-drop alerts, price history, stock alerts. It flags product compatibility (for example, when building a PC, it proposes alternatives if components are incompatible). Integrates with Google Wallet for loyalty info and merchant offers. Summer 2026 in the US.

Workspace — Live everywhere

The “Live” trend extends across the whole Workspace stack. Gmail Live enables conversational email search, AI Inbox intelligent inbox management. Docs Live lets you create and edit documents conversationally. Google Keep gets a similar mode that organises free-flowing thoughts into concise notes. Launch: summer 2026, US-first for AI Pro and Ultra subscribers (English only) on Android and iOS.

Google Pics is a new standalone app for AI image generation and design. Details, pricing, and timeline were not announced on stage.

On-stage demo — Gemini for macOS:the presenter selected several different files in Finder (invoices, notes, images) and with a voice command instructed Gemini to extract key details, organize them, and generate a polished Gmail email. The demo took a few seconds — a clear effort to show “native OS integration” similar to Apple Intelligence.

Android XR — Intelligent Eyewear

Google announced the first audio glasses in the Android XR platform, shipping fall 2026. Hardware is made by Samsung in collaboration with Qualcomm. External designs come from Gentle Monster and Warby Parker — a clear push to move hardware from tech niche into fashion mainstream. Pairing works with both Android phones and iPhone.

On-stage live demo:the presenter showcased two working models of glasses — one in the Warby Parker design, the other Gentle Monster. The demo included hands-free activation of Gemini by voice, contextual AI answers directly to the ear without pulling out a phone, and basic navigation. Pricing was not mentioned in the keynote.

Android for developers

• Android CLI (stable): enables AI agents to access Android Studio capabilities — SDK downloads, device testing.
• Android Skills (open-source): a set of workflows that help LLM models follow Android best practices. Includes migration skills for Jetpack Compose and Navigation 3.
• Android Bench: an LLM leaderboard for Android dev tasks. Includes open-weight models like Gemma 4. Available at d.android.com/bench.
• Migration Agent (Android Studio): previewed, not yet released. Migrates code into native Kotlin Android apps from React Native, web frameworks, or iOS source. Google claims it cuts weeks to hours.
• Android Halo: a new agent progress indicator at the top of the screen that visualises what your agent is working on. “Later 2026” for Gemini Spark and other supported agents.

Web platform for AI agents

WebMCP (Web Model Context Protocol):a newly proposed open web standard that allows developers to expose structured tools — JavaScript functions, HTML forms — to AI agents running in the browser. Goal: faster and more reliable task completion than DOM scraping. Launch: experimental origin trial in Chrome 149; Gemini in Chrome support soon.

Modern Web Guidance: a set of expert-vetted skills for coding agents, covering over 100 use-case scenarios. Integrates with Baseline. One-click install in Antigravity or via CLI command npx modern-web-guidance install.

Chrome DevTools for Agents:Chrome DevTools capabilities made available to agents — verification, debugging, code optimisation, automated quality audits, user experience emulation, hand-over session with auto-connect.

HTML-in-Canvas API: lets developers build immersive 3D experiences that remain searchable, accessible, and interactable. Integrates DOM elements into canvas with WebGL and WebGPU support. Currently in origin trial.

Antigravity 2.0, CLI, SDK, and Managed Agents

Google advanced its agent-first development platform to version 2.0, with new access paths for different developer profiles:

• Antigravity 2.0 — updated web IDE with better subagent orchestration, sandboxing, credential masking, and hardened Git policies.
• Antigravity CLI — terminal interface for dev teams used to CLI workflows.
• Antigravity SDK — programmatic access to the agent harness with the option to deploy on your own infrastructure.
• Managed Agents in the Gemini API — a single API call returns a fully provisioned agent with a remote sandbox. Removes infrastructure-setup friction.
• Google AI Studio received native Kotlin support for Android dev, Workspace integrations, one-click deploy to Cloud Run and Firebase. Projects can be exported seamlessly to Antigravity.

Security and content authenticity

In the last third of the keynote, Hassabis returned to the topic of safety in agentic systems. The following were mentioned:

• SynthID — Google's AI verification technology will be extended to Search and Chrome. Goal: detect AI-generated content directly in the browser.
• C2PA Content Credentials — a standard that lets you verify whether content is an unedited original from the camera or has been modified by tools.
• Code Mender — a new product that automatically scans a codebase, identifies security vulnerabilities, and proposes patches.

YouTube — Ask YouTube and Omni in Shorts

Ask YouTube handles complex search queries and follow-up questions, displaying results as an interactive structured response from the list of most relevant videos. Currently for Premium subscribers via youtube.com/new (US only).

Gemini Omni in YouTube Shorts: integrated into Remix and Create app for fast AI editing of short videos.

Conclusion — the real main message

Two big narratives come out of the two-hour keynote. First: Google is shifting its AI strategy from “assistant” to “agent.” Spark, Information Agents, Universal Cart, Android Halo — all these products assume AI will work in the background, initiate actions on its own, and communicate results only when relevant. This is a significant shift from today's “ChatGPT-style” one-prompt / one-response model.

Second: Google wants to play a double game. Consumer products (Gemini app, Search AI Mode, Omni video) are free or cheap inside the ecosystem to build a defensible distribution layer. Dev and enterprise products (Antigravity SDK, Managed Agents, WebMCP) are technically more ambitious and deliberately open to win developer loyalty. That is a different strategy from OpenAI (consumer-first via ChatGPT) and Anthropic (enterprise-first via the Claude API).

And a third, quieter one: 85+ sessions, codelabs, and additional resources are available on-demand from May 21, 2026 — so the announcements that did not fit into the keynote spread out over weeks. Google is evidently calculating that the pace of AI announcements can no longer be compressed into two hours. I/O 2026 was more a “curated introduction” than a “complete release log.”

Frequently asked questions

When exactly was Google I/O 2026?

The main keynote was on May 19, 2026, at the Shoreline Amphitheatre in Mountain View. 85+ further sessions and codelabs are available on-demand from May 21, 2026.

When will Gemini 3.5 Pro be available?

According to the keynote, “next month” — June 2026. Currently in testing with selected partners. Details on pricing, context window, and benchmarks were left for a separate announcement.

Who makes the hardware for Android XR Intelligent Eyewear?

Hardware is made by Samsung and Qualcomm. External design is delivered by Gentle Monster and Warby Parker. Pairing is supported by both Android phones and iPhone. Launch: fall 2026.

What is the difference between Gemini Spark and Information Agents?

Gemini Sparkis a personal agent integrated with your Workspace data — it handles planning and communication tasks (wedding, emails, documents). Information Agents monitor external web sources (blogs, news, financial data) and alert you to changes. Spark = internal worker, Information Agent = external scout.

What changed in Google AI Ultra subscriptions?

Three changes: (1) new starting tier $100/month with 5× higher limits than AI Pro; (2) the previous $250 plan reduced to $200, same capabilities; (3) switch to a compute-used model — limits are metered by prompt complexity and features used, not by the number of daily prompts. Limits refresh every 5 hours up to a weekly cap.

Sources

This recap was assembled from the following official and editorial sources published May 19–20, 2026: Google Blog (Sundar Pichai's opening keynote), Google Developers Blog, 9to5Google, Tom's Guide, Engadget, TechRadar, Digit.in.

Note: this article is an editorial recap. Sundar Pichai's quotes are recorded in the official Google blog transcript. Details on Gemini 3.5 Flash API pricing come from third-party sources and should be verified against the official Google AI for Developers documentation before publishing any specific numbers.