Google I/O 2026: Gemini 3.5 Flash, Antigravity, Universal Cart, Gemma 4

Google I/O 2026 delivered Gemini 3.5 Flash — 4x faster than competing frontier models at 289 tokens per second — alongside Antigravity, an agent-first IDE going directly at Cursor and Claude Code. That combination, plus Googlebooks (the Chromebook replacement running a unified Android+ChromeOS platform), Android XR glasses with Warby Parker and Gucci partnerships, and a Gemini user base that has doubled to 900 million monthly actives, makes this the most developer-relevant Google I/O in years.

Here is everything announced at the May 19 keynote, with the developer and infrastructure implications for each.

Gemini 3.5 Flash: The Speed Play

Gemini 3.5 Flash is the headline model release. The numbers: 289 tokens per second output speed, 4x faster than comparable frontier models. It outperforms Gemini 3.1 Pro on almost all benchmarks while being significantly cheaper.

Google has not published official pricing yet, but industry estimates put it at approximately $1.50 per million input tokens and $9 per million output tokens — roughly 3x higher than Gemini 3 Flash. For comparison: Claude Sonnet 4.6 runs $3/$15, GPT-4o runs $5/$15.

The model is optimised specifically for agentic tasks, coding, and long-horizon reasoning. Available now in AI Studio and the Gemini API. The speed advantage over Anthropic and OpenAI's comparable tier is meaningful for high-throughput applications — streaming responses, real-time agent loops, customer-facing latency-sensitive deployments.

Antigravity 2.0: 12x Faster, 16 Specialized Agents

Google launched Antigravity 2.0, an agent-first IDE powered by Gemini 3.5 Flash. The desktop app, CLI, and SDK are all available globally now.

The headline number: Antigravity 2.0 is 12x faster than the previous version. Token usage is kept down by design — the architecture avoids burning context on boilerplate operations.

The core upgrade is AgentKit 2.0 — a framework of 16 specialized agents that Antigravity deploys in parallel across a codebase:

• Frontend design, backend logic, security review, automated testing, infrastructure provisioning, SEO, mobile layout, UX, database configuration
• Each agent comes pre-loaded with 40+ domain-specific skills
• Agents coordinate on tasks rather than running sequentially

Managed Agents via Gemini API: Google simultaneously announced that developers can now spin up Gemini agents with a single API call. These agents reason, use tools, and execute code in isolated Linux environments — a direct competitor to Anthropic's Claude Code and OpenAI's Codex agent APIs.

The competitive context: Cursor has approximately 1 million paid users as of early 2026. Claude Code is Anthropic's coding agent. GitHub Copilot holds enterprise incumbent position. Antigravity enters with 12x speed, Google infrastructure, and native integration with Google Cloud, BigQuery, Workspace, and Firebase. Pricing not announced yet.

Universal Commerce Protocol: Google Rewires Shopping

Vidya Srinivasan, Google's VP of Ads and Commerce, announced two connected products that reframe how AI agents interact with the commercial web.

Universal Commerce Protocol (UCP) is an open-source protocol that gives AI agents and merchant systems a common language across the entire shopping journey — discovery, comparison, validation, checkout. Google is positioning UCP as an industry standard. Founding partners: Amazon, Walmart, Best Buy, Microsoft, Salesforce, and Stripe. The protocol is open-source; any merchant or agent builder can implement it.

Universal Cart is the consumer product built on UCP. It is an intelligent shopping cart inside Google Search that:

• Automatically looks up price history and current price changes across retailers
• Validates product compatibility before purchase (the demo example: detecting an incompatible motherboard and processor combination before the user completes the order)
• Allows checkout via Google Pay directly or transfers the cart to the retailer's site
• Enables AI agents — including Gemini Spark — to make purchases autonomously when the user authorises it

The combination is significant. Universal Cart turns Google Search into a transaction layer, not just a discovery layer. The founding partner list (Amazon and Walmart as competitors to each other, Microsoft and Stripe as infrastructure) means Google has secured enough supply-side coverage at launch to make the protocol immediately useful.

For developers building commerce applications: UCP is the protocol to watch. If it achieves the adoption Google is targeting, any agent needing to transact in the real world will speak UCP.

Googlebooks: The Chromebook Replacement

Chromebooks are being replaced by Googlebooks. The new platform runs Aluminum OS — a unified operating system that merges Android and ChromeOS into a single platform with a single development target.

Key features: Magic Pointer (wiggle cursor to trigger AI-assisted image comparison and manipulation), Android phone mirroring and file browsing, Gemini built-in, full Android app support without virtualisation. Cross-Device Quick Share now interoperates with AirDrop on compatible phones.

Hardware partners at launch: Acer, ASUS, Dell, HP, Lenovo. First devices arrive fall 2026. Pricing not announced.

The developer implication is significant. Aluminum OS unifies app deployment across Android's 3.2 billion devices and Googlebooks. If you build Android apps, they run on Googlebooks without modification. The Gemini integration in Aluminum OS opens native AI APIs at the OS layer — not a browser plugin or third-party library but a first-class system API.

Android XR Glasses

Android XR glasses are going to market with a fashion-first strategy that inverts the Google Glass mistake. Hardware partners: Samsung (Galaxy Glasses), XREAL (Project Aura), and fashion labels Warby Parker, Gentle Monster, and Gucci (2027 launch from Kering).

Two product tiers:

AI glasses (display-free): Cameras, speakers, microphones for hands-free Gemini Live interaction. Real-time translation, navigation, messaging, visual understanding. No in-lens display. These compete with Meta Ray-Bans directly.

Glasses with in-lens display: All of the above plus private contextual overlays visible only to the wearer. The killer use case Google is demonstrating: translated subtitles from a conversation in another language appearing in your field of view.

Samsung Galaxy Glasses confirmed specs: Qualcomm Snapdragon AR1 chipset, 155mAh battery, 12MP Sony camera, photochromic lenses, approximately 50g. XREAL Project Aura uses an external battery puck for lighter frame weight.

For developers: the Android XR SDK is the platform target. If you've built for Meta Quest or Apple Vision Pro, Android XR is a third platform. Samsung's distribution scale gives it volume potential the moment consumer glasses ship in 2026.

Android 17: Intelligence as the OS Layer

Android 17 frames Gemini as the core intelligence layer of the OS rather than an add-on application. Developer-relevant additions:

Create My Widget: Natural language widget creation. Users describe what they want; Gemini builds a home screen widget. This opens a new surface — AI-generated UI components sitting on the same home screen as developer-crafted widgets.

Cross-Device Quick Share with AirDrop interop: Rolling out to Pixel, Samsung, OPPO, OnePlus, Vivo, Xiaomi, and HONOR throughout 2026. Direct answer to Apple's ecosystem lock-in.

Gboard Rambler: Voice-to-text with Gemini cleanup — transcribes, corrects, and formats spoken input. Built into the keyboard, available to all apps with no developer integration required.

Pause Point: 10-second pause screen for distracting apps — platform-level screen time management baked into the OS.

Enhanced Privacy: Granular per-app location control. Improved Scam Protection: AI-powered call and message analysis at the system level.

Gemini Omni: Video Generation

Gemini Omni is a multimodal video generation model arriving this summer. Input: audio, video, images, or text in any combination. Output: up to approximately 10 seconds per clip with synchronised audio — lip sync, footsteps, ambient sound. Physics simulation accuracy is a stated capability.

The in-chat editing workflow allows iterative clip refinement without leaving the conversation. This positions Gemini Omni as a direct consumer competitor to OpenAI's Sora and puts video generation inside the 900 million-user Gemini app rather than a separate product.

Google AI Ultra at $100/Month and Gemini Spark

Google announced Google AI Ultra — a new premium subscription tier starting at $100 per month. This is Google's direct answer to OpenAI's ChatGPT Pro ($200/month) and Anthropic's Claude Max ($100/month), positioned as the highest-capability consumer subscription Google offers. Google AI Pro, a lower tier at $19.99/month, is planned to launch within 90 days for broader access.

The headline exclusive inside AI Ultra is Gemini Spark — not a new model, but a persistent autonomous AI agent. Spark is the consumer launch of what Google was codenamed "Remy" internally. It ships to Google AI Ultra subscribers in the US next week.

What makes Spark different from regular Gemini: it does not wait for prompts. Regular Gemini is a session-based chatbot — you ask, it answers. Spark is a persistent, always-on agent that monitors your Gmail, Google Calendar, Google Docs, and Google Sheets, and takes action proactively without you explicitly requesting it. It runs in the background 24/7, learns from your patterns — tools you use, decisions you repeat — and acts on recurring tasks automatically.

The demo at the keynote showed a concrete example: a user asks Spark to help organise a party. Spark creates an RSVP tracker in Google Sheets, automatically updates it as people respond via Gmail, and sends follow-up reminders to non-responders — all without additional user input. The workflow spans three Google products and runs asynchronously.

The Gemini app gets a new two-tab interface to support Spark: Chat (current session-based Gemini) and Agent (where Spark tasks are created, monitored, and scheduled). Other Spark capabilities shown: automated email summaries and archive, newsletter unsubscribe automation, calendar meeting briefs generated before important events, personalised news digests that track ongoing stories, and multi-step web-based workflows.

One important caveat Google announced explicitly: Spark is marked experimental. Google warned that it may share your information with third parties or make purchases on your behalf without always asking for permission first. Active supervision is recommended. It is not designed for medical, legal, or financial advice.

The pricing comparison is significant for anyone evaluating premium AI subscriptions: Google AI Ultra at $100/month equals Claude Max and costs half of ChatGPT Pro. The $19.99 Google AI Pro tier arriving within 90 days will give Spark access to a much larger user base.

Android Halo and the Agent Dashboard

Android 17 ships a new surface called Android Halo — a dedicated homebase for all AI agents running on the device. Halo shows at-a-glance visibility of what every agent is currently doing, communicates agent progress through a subtle indicator at the top of the screen, and lets users monitor and manage active tasks without opening each app separately.

Halo is the interface layer that makes Gemini Spark — and any third-party agent built on Android XR — legible to users. Without it, autonomous agents running in the background are invisible until they complete a task. Halo surfaces them as they work. Availability: later this year, launching first for Gemini Spark and supported agents.

Gemma 4: Open-Source Release

Google released Gemma 4, the latest generation of its open-source model family, at I/O 2026. Apache 2.0 licence — commercially permissive.

Four model sizes:

• Effective 2B (E2B) and Effective 4B (E4B): Both support native audio input for speech recognition
• 26B Mixture of Experts (MoE): Sixth place on Arena AI text leaderboard
• 31B Dense: Third place on Arena AI text leaderboard

All four models support native function calling, structured JSON output, system instructions, video and image processing at variable resolutions.

The positioning: Google claims Gemma 4 models outcompete rivals up to 20x their size. The Gemma family has now been downloaded 400 million times, with over 100,000 community variants built on top of the model weights.

For developers self-hosting or fine-tuning: the 31B dense model at third place on Arena is a meaningful open-weight option at a size that runs on high-end consumer hardware. The 26B MoE trades some absolute quality for efficiency.

Other Announcements

Gemini 3.5 Pro: Currently in testing, releasing next month. Higher capability than Gemini 3.5 Flash, positioned for complex reasoning tasks.

AI Mode in Search: Now has 1 billion monthly users. Queries doubled every quarter since launch. Upgraded to Gemini 3.5 Flash as the default model.

Intelligent Search Box: Biggest redesign of the Google Search input in 25+ years. Dynamically expands for longer queries, supports multi-modal input (text, images, files, videos, Chrome tabs), with AI-powered suggestions that go beyond autocomplete. Rolling out today in all countries where AI Mode is available.

Daily Brief in Gmail: Personalized daily digest generated from Gmail, Calendar, and Tasks, with prioritised to-dos and suggested next steps. Available today in the US for Google AI Pro and Ultra subscribers.

Docs Live: Voice-to-document generation in Google Docs, Sheets, and Gmail. Natural language editing and formatting commands. Summer 2026.

Ask YouTube: AI-summarised video search with step-by-step extraction from video content. Rolling out across the US this summer.

Circle to Search + SynthID: AI-generated image detection integrated into Chrome and Search. Verify whether images are AI-generated or manipulated. Rolling out this summer.

8i TPU: New tensor processor for distributed training across global infrastructure — the hardware layer behind 3.2 quadrillion tokens processed per month.

"AGI Is Now on the Horizon" — Demis Hassabis

The most-quoted line from the entire keynote came from Demis Hassabis, CEO of Google DeepMind: "AGI is now on the horizon." He made the statement in the context of emphasising the safety requirements for agentic AI systems — the same systems being deployed through Gemini Spark and the Android 17 agent layer.

This is not a throwaway line. Hassabis is one of the most technically credible figures in AI research. His saying AGI is on the horizon at a public keynote, rather than in a research paper, is a deliberate public positioning statement. It lands the same week Anthropic published safety research on frontier models and OpenAI raised its largest funding round. The frontier labs are converging on a shared claim: the systems being built now are qualitatively different from anything before.

Jules: Google's Autonomous Coding Agent

Jules is Google's asynchronous coding agent, announced as generally available at I/O 2026. It does not work alongside you — it works instead of you for discrete tasks.

The workflow: assign Jules a complete task (fix a bug, implement a feature, refactor a module). Jules runs the task in an isolated virtual machine, produces the code changes, and returns a pull request ready for review. No back-and-forth, no prompt engineering mid-task. Jules uses Gemini 3.1 Pro on paid plans and Gemini 3 Flash on the free tier.

Jules is the clearest demonstration of what Google means by "agentic AI" in practice — not a copilot that suggests the next line, but an agent that owns a task end-to-end. It competes directly with GitHub Copilot Workspace and Anthropic's Claude Code agentic workflows.

More Announcements from the Developer Sessions

Code Mender: Google's automated security tool. It scans codebases for vulnerabilities and patches them automatically. No manual triage. Announced as a surprise addition — not previewed before I/O.

Mariner: Google's web browsing agent, announced in preview. Mariner can navigate the web autonomously to complete research and transaction tasks — the same capability class as OpenAI's Operator.

Wireless iPhone-to-Android transfer: Android 17 ships a wireless device switching tool that works with iPhones. This is the most direct feature Google has ever built targeting Apple ecosystem users. No cable, no Google account required on the iPhone side.

Wear OS 7: New widget system, live update tiles, and Gemini Intelligence integration on the wrist. Rolling out with Android 17.

The Scale Numbers

Two metrics define Google's AI position heading into the second half of 2026:

900 million monthly active Gemini users — doubled in the past year. For context, ChatGPT reported approximately 500 million weekly users in early 2026. Gemini's monthly active count now materially exceeds OpenAI's comparable metric.

3.2 quadrillion tokens processed per month — 7x year-over-year growth. This is the infrastructure number: Google's TPU fleet is handling token volumes that require continuous capacity expansion, which explains the 8i TPU announcement and the aggressive Gemini API pricing strategy.

13 Google products have passed 1 billion users. 5 have passed 3 billion users. Gemini is now embedded in the distribution stack of every one of them.

Key Takeaways

• Gemini 3.5 Flash: 289 tokens/second, 4x faster than competing frontier models, outperforms Gemini 3.1 Pro on most benchmarks; estimated ~$1.50/$9 per million tokens; available now in AI Studio and Gemini API
• Antigravity 2.0: Agent-first IDE, 12x faster than previous version; AgentKit 2.0 with 16 specialised agents (40+ skills each); desktop app + CLI + SDK available globally now; Managed Agents Gemini API lets developers spin up agents with single API call
• Universal Commerce Protocol (UCP): Open-source protocol for AI agents to transact across the web; founding partners Amazon, Walmart, Best Buy, Microsoft, Salesforce, Stripe; Universal Cart in Search validates product compatibility, pulls price history, enables agent-driven purchases
• Googlebooks: Chromebook successor running Aluminum OS (Android + ChromeOS unified); hardware partners Acer, ASUS, Dell, HP, Lenovo; fall 2026; single development target for Android + laptop
• Android XR Glasses: Samsung Galaxy Glasses (Snapdragon AR1, 12MP Sony camera, ~50g) + XREAL Project Aura + Warby Parker, Gentle Monster, Gucci (2027); display-free and in-lens display tiers; Gemini 2.5 Pro
• Android 17: Create My Widget (AI-generated home screen widgets), Cross-Device Quick Share with AirDrop interop (6 OEMs rolling out 2026), Gboard Rambler (Gemini voice cleanup), Pause Point (screen time)
• Gemini Omni: Multimodal video generation; up to 10 seconds per clip; synchronised audio; in-chat editing; summer 2026
• Google AI Ultra: New $100/month premium plan — matches Claude Max ($100/mo), half of ChatGPT Pro ($200/mo); Google AI Pro ($19.99/mo) arriving within 90 days
• Gemini Spark: Persistent 24/7 autonomous agent (codenamed "Remy"), exclusive to AI Ultra; proactive task execution across Gmail, Sheets, Calendar without prompts; live demo showed party RSVP automation across 3 Google products; experimental — may make purchases or share data without asking; available next week in US
• Android Halo: New agent homebase showing all active agents at a glance; subtle top-screen progress indicator; launching later this year with Gemini Spark
• Gemma 4: Open-source release (Apache 2.0); four sizes — E2B, E4B, 26B MoE (6th Arena), 31B Dense (3rd Arena); 400M total downloads, 100K+ community variants; native audio, function calling, multimodal
• Jules GA: Google's async coding agent now generally available; receives task → runs in isolated VM → returns pull request; Gemini 3.1 Pro on paid, Gemini 3 Flash on free
• Code Mender: Automated vulnerability scanning and patching for codebases; announced as surprise addition
• Mariner: Web browsing agent in preview; autonomous web navigation for research and transactions
• "AGI is now on the horizon": Demis Hassabis (Google DeepMind CEO) at the keynote — most-cited quote from the event
• Gemini 3.5 Pro: In testing now, releasing next month
• AI Mode in Search: 1 billion monthly users; queries doubled every quarter since launch; Gemini 3.5 Flash now the default model
• Scale: 900M Gemini MAUs (2x YoY), 3.2 quadrillion tokens/month (7x YoY), 13 Google products at 1B+ users, 5 at 3B+ users

For the preview published before the keynote, read Google I/O 2026: Android 17, Gemini Intelligence, Googlebooks, XR Preview. For the Anthropic and PwC enterprise AI deployment that Antigravity now competes against, read Anthropic + PwC: 30,000 on Claude, Insurance Drops to 10 Days.