Google I/O 2026: Complete AI Announcement Guide

Google I/O 2026 opened this morning at Shoreline Amphitheatre with Sundar Pichai delivering more than 100 AI announcements — Gemini 3.5 Flash going generally available, Antigravity 2.0 launching as a standalone agent platform, a new 24/7 personal agent called Gemini Spark, a multimodal generative family called Omni, AI Mode in Search crossing 1 billion monthly users, and a repriced AI Ultra subscription dropping from $250 to $200/mo with a new $100/mo developer entry tier.

This is the single biggest announcement day in the May 16–24 vertical-integration wave — a window that also included Cursor Composer 2.5 (May 18), a Blackstone × Google $5B TPU joint venture (May 18), and SAP Sapphire Madrid (May 19). Anthropic counter-programmed with Code with Claude London opening the same morning, shipping self-hosted sandboxes and MCP tunnels in parallel. Two products named "Managed Agents" shipped from Google and Anthropic on the same day with opposite architectures.

This hub post indexes every major I/O 2026 announcement, links each section to the authoritative deep-dive in our archive, and flags the 11 canonical reconciliations that matter most for accurate reporting. Use the table of contents above to jump to any announcement category, or read straight through for the full picture.

Sundar Pichai opened the I/O 2026 keynote with six scale numbers that frame the agent era as already arrived, not approaching. Across Google products, the company now processes 3.2 quadrillion tokens monthly — 7× year-over-year growth. The Gemini app has reached 900 million monthly active users (up from 400M at the last earnings call). AI Overviews in Search serves 2.5 billion monthly active users. AI Mode in Search, one year after launch, has crossed 1 billion monthly active users with queries more than doubling every quarter. Eight and a half million developers build with Google models monthly. SynthID has watermarked more than 100 billion images across Google products.

Pichai's framing was explicit: "We're now in the part of the AI cycle where people want to see the value in the products they use every day." The keynote was structured as a product ledger — every number anchored to a product surface, every product surface with a deep-dive announced immediately after.

What the numbers don't show: the day before I/O, Blackstone and Google announced a $5B TPU joint venture ("N1") targeting 500 MW online in 2027. Cursor launched Composer 2.5 the same day. I/O sits inside a 48-hour AI-infrastructure compression that most keynote coverage will treat in isolation. For the multi-vendor context, see our AI coding IDE landscape as of May 2026.

Gemini 3.5 Flash launched generally available today with stable model ID gemini-3.5-flash — replacing gemini-3-flash-preview as the default model in the Gemini app and AI Mode in Google Search worldwide. This is the centerpiece of the keynote and the most consequential model launch of the day.

The Flash-first cadence is the strategic story. Google's historical pattern was Pro first, Flash later — announcing the full-capability model at I/O and rolling out the efficient sibling months after. I/O 2026 inverts this: Gemini 3.5 Flash ships today; Gemini 3.5 Pro rolls out in June 2026. Pichai was direct about why — Antigravity 2.0 was built using Gemini 3.5 Flash, and the speed/cost profile of Flash is what makes the agent harness economically viable at consumer scale.

Pricing for the Gemini API is reportedly $1.50 per million tokens input and $9.00 per million tokens output (standard tier) — roughly 25% cheaper than Gemini 3.1 Pro and approximately 3.3× cheaper on input than the Pro tier of GPT-5.5. For benchmark details, API specs, and the thinking_level parameter replacing thinking_budget, see our already-published authoritative deep-dives: Gemini 3.5 Flash: Benchmarks, Thinking & API Guide and the head-to-head Gemini 3.5 Flash vs GPT-5.5 vs Opus 4.7 on agentic coding. This hub post does not republish the benchmark table or API details — those live in the linked posts.

Antigravity 2.0 launched today as a native standalone desktop application built around multi-agent orchestration — alongside an Antigravity CLI, Antigravity SDK, and Managed Agents in the Gemini API. This is a significant reframe: Antigravity 1.0 (November 18, 2025) was a VS Code fork with Manager View, five parallel agents, and Chrome integration. Antigravity 2.0 is a standalone agent-first platform, not an IDE.

Pichai cited Antigravity 2.0 from the keynote stage in an important direction: Google used Antigravity 2.0 to build Gemini 3.5 Flash itself — the first time the company has publicly demonstrated self-hosting of its own AI development on its agentic IDE. The 2.0 platform ships with native voice command support, integrations with Google AI Studio, Android Studio, and Firebase, plus scheduled task execution and custom subagent workflows.

The Antigravity CLI and SDK are available now. The enterprise integration via Gemini Enterprise Agent Platform rolls out over the coming months. Critically: the Gemini CLI is being sunset June 18, 2026 — Pro and Ultra customers must migrate to Antigravity 2.0 before that date. For the full feature matrix, CLI commands, SDK patterns, AI Ultra gating, and the rocky five-day rollout story, see the Antigravity 2.0 deep dive — desktop app, CLI, SDK, and the AI Ultra gating.

Gemini Spark is Google's 24/7 personal AI agent inside the Gemini app, powered by Gemini 3.5 Flash and the Antigravity harness so it can run long-horizon background tasks while you're offline. Where Antigravity 2.0 is the developer surface — desktop app, CLI, SDK — Spark is the consumer reference implementation on the same underlying agent harness.

Spark works across Gmail, Calendar, Drive, YouTube, Maps, and third-party apps including Canva, OpenTable, and Instacart. The integration list is expected to expand significantly over the summer. One critical clarification: the Agent Payments Protocol 2 (AP2) is NOT in Universal Cart at launch. AP2 ships separately with Gemini Spark "in coming months" — the Universal Cart announcement uses Merchant Center integration only.

Availability: trusted testers receive access today (May 19). The Beta opens to AI Ultra subscribers in the U.S. the week of May 25. AI Pro subscribers follow in a subsequent wave. Spark is not yet globally available. For the full integrations list, availability timeline, and the Daily Brief feature that anchors Spark's ambient assistant story, see Gemini Spark, Google's 24/7 personal agent — full integrations list.

Gemini Omni Flash launched today as the first model in a new "Omni" family — generating output in any modality from any input (text, image, video, audio). The headline use case is video generation: Omni Flash can generate short video clips from text, image, or audio input, with a cap (reportedly 10 seconds for the initial release). All Omni output is watermarked with SynthID, with verification rolling into the Gemini app, Search, and Chrome in the coming weeks.

Access is tiered: Omni Flash is free for YouTube users via YouTube Shorts Remix, and gated to AI Plus, Pro, and Ultra subscribers in the Gemini app and Google Flow. Voice and speech editing capabilities were previewed but are not shipping in the initial Omni release — that functionality is deferred to a later update.

The election-year hedge is notable: Google has added explicit SynthID watermarking to all Omni outputs and is rolling out verification tooling in Chrome before the November 2026 election cycle. For the full Omni Flash guide covering text-to-video, image-to-video, audio-to-video, and SynthID verification, see Gemini Omni Flash — text→video, image→video, audio→video full guide.

The two-surface scoreboard at I/O 2026: AI Mode in Search has surpassed 1 billion monthly active users with queries more than doubling every quarter since launch. AI Overviews serves 2.5 billion monthly active users. Together, these numbers represent Google's clearest public framing of how deeply AI has restructured the Search product — two distinct surfaces, one embedded (Overviews) and one modal (AI Mode), each with scale that no competitor has publicly matched.

The Search box itself gets what Google calls "the biggest upgrade in over 25 years." The new box accepts text, images, files, videos, and Chrome tabs as input — and includes "intent anticipation," AI-powered suggestions that go beyond autocomplete to predict what a user needs before they finish typing. Information Agents — web-monitoring agents that run 24/7 in the background — roll out summer 2026, gated to AI Pro and AI Ultra subscribers.

SEO implications are significant. The May 21 Core Update (forecast for Wednesday, two days after I/O) lands the same week as the Search box redesign rollout — a potential algorithmic-plus-UI double-whammy. For the full SEO impact analysis, click-share modeling, and the Information Agents timeline, see Google Search overhaul: AI Mode at 1B monthly users — SEO impact analysis.

Universal Cart launched at I/O 2026 as an AI-powered cross-merchant shopping hub spanning Search, the Gemini app, YouTube, and Gmail — with price-history tracking, restock alerts, and AI compatibility checks (for example, it can analyze a custom PC build across multiple retailers simultaneously to verify component compatibility). Launch partners include Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair, and Shopify merchants Fenty and Steve Madden.

Key clarifications: Universal Cart rolls out across Search and the Gemini app in the U.S. during summer 2026; YouTube and Gmail integrations follow later. It is emphatically not a global launch. The Agent Payments Protocol 2 (AP2) is NOT part of Universal Cart at launch — AP2 ships separately with Gemini Spark "in coming months." Universal Cart uses Merchant Center integration only at launch. Merchants should update their product feeds with the native_commerce attribute to signal Universal Cart eligibility.

For the merchant preparation guide, feed requirements, AP2 timeline, and the Universal Commerce Protocol (UCP) context that Universal Cart rides on, see Universal Cart merchant preparation guide.

Google's Managed Agents in the Gemini API launched today — a single API call provisions a remote Linux execution environment so an agent can reason, execute code in a sandbox, and browse the web without the developer managing infrastructure. The Interactions API is the new developer surface for managed agents, accessible via Google AI Studio and the Gemini Enterprise Agent Platform.

The name collision is the most interesting framing of the day: Google ships "Managed Agents API" and Anthropic ships "Claude Managed Agents" on the same day, with opposite architectures. Google's managed agents are hosted — Google provisions the execution environment. Anthropic's managed agents are self-hosted — enterprise customers run Claude agents inside their own VPC with Cloudflare, Daytona, Modal, or Vercel providing the sandbox. The architectural divergence is intentional and signals a genuine strategic difference in how the two companies view enterprise security models.

For the full Managed Agents API spec — single-call patterns, the Interactions API surface, sandboxing behavior, and the Google-vs-Anthropic architecture comparison — see Google Managed Agents API: agents in a single call.

The AI Ultra subscription drops from $250 to $200/mo at I/O 2026 — 20× the usage limits of AI Pro, full access to Gemini Spark in beta, Antigravity 2.0 quota priority, and YouTube Premium. A new $100/mo tier (positioned as developer/technical worker) provides 5× Pro tier limits and 20 TB Drive storage. This is a quiet but material pricing move: $200/mo AI Ultra now competes head-on with ChatGPT Pro ($200) and the Claude Max plan.

AI Pro ($19.99/mo) now includes YouTube Premium Lite at no extra charge (an $8.99/mo standalone value). Information Agents — the web-monitoring search agents that run 24/7 — are gated to AI Pro and AI Ultra (not Ultra-only). The AI Plus tier ($7.99/mo) retains access to Gemini Omni Flash via the Gemini app and Google Flow.

Code with Claude London opened this morning — Anthropic's first European developer summit, running May 19–20 at Park Plaza Westminster Bridge — directly overlapping Google I/O. The counter-programming framing is about timing, not direct technical response: both products were planned independently. The collision is informative regardless.

Anthropic shipped two products at the London event: self-hosted sandboxes in public beta and MCP tunnels in research preview. Self-hosted sandboxes let enterprise customers run Claude agents inside their own VPC — the sandbox where the agent executes tools and the services it reaches both run within the established boundaries of the enterprise, under the customer's security and runtime controls. Day-one sandbox providers announced: Cloudflare (with Amplitude as the reference enterprise customer), Daytona (Clay), Modal (DoorDash), and Vercel (Rogo). MCP tunnels let agents reach MCP servers inside a private network without exposing those servers to the public internet via a lightweight outbound-only gateway.

A notable adjacent signal: Andrej Karpathy joined Anthropic on May 19, 2026 — a research-talent event on an already-crowded day. Anthropic also re-asserted the February 4, 2026 ad-free essay at the London event — framing the self-hosted architecture as consistent with the privacy-first, no-ads positioning announced earlier in the year.

For the full Code with Claude London recap — sandboxes, MCP tunnels, the Karpathy signal, and the Google vs Anthropic managed-agent architecture comparison — see Code with Claude London: self-hosted sandboxes + MCP tunnels launch.

Not everything announced at I/O 2026 shipped today. Several products and features were previewed or confirmed with delayed timelines. The most consequential: Gemini 3.5 Pro is confirmed for June 2026 — it is not available today and should not be framed as launched. The Flash-first cadence means Pro follows next month.

Universal Cart's YouTube and Gmail surfaces are deferred — launch covers Search and the Gemini app only (U.S., summer 2026). Information Agents (web-monitoring search agents for AI Pro and AI Ultra) roll out summer 2026, not today. AP2 (Agent Payments Protocol 2) ships separately with Gemini Spark "in coming months" — AP2 is not part of Universal Cart at launch.

Gemini Omni voice and speech editing were previewed but not shipped in the initial Omni Flash release. Android XR Intelligent Eyewear audio glasses (Gentle Monster, Warby Parker, Samsung partnership) ship fall 2026 — today was announcement only. The Gemini CLI sunset (June 18, 2026) is a deadline, not a feature — Pro and Ultra customers need to migrate to Antigravity 2.0 before that date.

The I/O news cycle doesn't end today. Five dates are live on the calendar for the remainder of the week and next month.

Wednesday May 21 — the Core Update is forecast to land two days after I/O. The timing creates a potential double-whammy for SEOs: a Google algorithmic update landing the same week as the Search box redesign rollout. If you're monitoring Search Console and seeing ranking volatility, separate the Core Update signal from the UI change before drawing conclusions. The AI Transformation implications for your organic channel are worth modeling this week — see our AI transformation services for structured SEO strategy in an AI Mode world.

Friday May 22 — Google resets global weekly Gemini quotas for all AI Pro and AI Ultra subscribers affected by the Antigravity 2.0 rollout instability (the "launched and apologized" five-day window). If your Antigravity 2.0 or Gemini app usage was throttled May 19–21, that reset applies.

Week of May 25 — Gemini Spark Beta opens to AI Ultra subscribers in the U.S. This is the first general-availability wave for Spark; trusted tester feedback from this week will shape the AI Pro rollout timeline. Watch the Spark integrations list closely — the Canva, OpenTable, and Instacart partnerships are confirmed; more third-party integrations are expected to be announced in the week after I/O.

June 2026 — Gemini 3.5 Pro rolls out. June 18 specifically — Gemini CLI sunset deadline. Developers still running CLI-based workflows need a migration plan before then.