Google I/O 2026 recap

Google shipped 16 separate announcements at I/O 2026 on May 19, and they all rhyme. Gemini 3.5 Flash is "frontier intelligence with action." The Gemini app added Spark, a 24/7 personal agent, and Daily Brief. Antigravity 2.0 dropped the IDE entirely and reorganized around agents that schedule themselves. Universal Cart turned Google Shopping into something agents can buy from on your behalf. The phrase Google kept using was "agents that help us act."

The action layer is the product now

A year ago, the question every AI assistant tried to answer was "what should I do." This year, Google's answer to that question is "I'll do it." Gemini 3.5 Flash is benchmarked on Terminal-Bench 2.1 (76.2%) and MCP Atlas (83.6%), both agent benchmarks, and Google leads with that framing instead of reasoning scores. Antigravity 2.0 ships as a standalone desktop app with no IDE, where the primary interaction is dispatching agents that can run on a schedule, spawn dynamic subagents, and asynchronously report back. Gemini Spark is the consumer version of the same idea: a thing that runs in the background and surfaces work, not a thing you open to ask a question.

The supporting cast tells the same story. Daily Brief is an agent that decides what you need to see before you've asked. Universal Cart is the infrastructure layer for shopping agents that complete transactions. Android Halo puts the agent in the status bar. Intelligent eyewear is the agent without a screen at all. Even the science work, Co-Scientist and the Computational Discovery prototype built on AlphaEvolve, is framed around agents that generate and test hypotheses in parallel rather than as a smarter search box.

Token volume is the clearest tell. Sundar Pichai said Google is now processing more than 3.2 quadrillion tokens per month, up from 480 trillion at I/O 2025. A 7x jump in twelve months is not answer-the-question usage. That's what running long-horizon work in the background looks like.

OpenAI and Anthropic showed up too

The same day Google ran I/O, OpenAI's only public release was a content provenance update adopting Google DeepMind's SynthID watermarking for ChatGPT, Codex, and API images, alongside C2PA conformance. Google's SynthID partnership announcement at I/O named OpenAI, Nvidia, ElevenLabs, and Kakao as integrators. The biggest news from OpenAI on Google's biggest day was that OpenAI agreed to use Google's watermark.

Anthropic counter-programmed differently. They held Code with Claude London, their first developer conference outside the US, and shipped self-hosted sandboxes for Claude Managed Agents alongside MCP tunnels in research preview. The framing was the inverse of Google's: keep the agent loop on Anthropic's infrastructure, but let enterprises run tool execution and private service access inside their own perimeter, with Cloudflare, Daytona, Modal, and Vercel as launch partners. Same category, opposite trust posture.

Three labs, one beat, three bets on where the trust boundary lives. Google wants you to use Gemini for everything. OpenAI is happy to share infrastructure when it's the standard. Anthropic is happy to give up the data center if it means landing the enterprise. The model wars are mostly settled at the top. The next year is about which of those bets the buyers actually want.

Track what each lab ships next on Google, OpenAI, and Anthropic.