Google I/O 2026: Gemini Becomes an Agent Platform, and Search Gets Its Biggest Rework in 25 Years

Google I/O 2026 wrapped on May 20. If you want the one-line summary, Forbes had it: Google turned Gemini into an agent platform. Pichai called this the start of the “agentic Gemini era” on stage, and over two days the word agent got attached to nearly every product shown.

Start with the numbers Pichai gave, because they explain why Google is comfortable betting this hard:

• The Gemini app has passed 900 million monthly users, up from 400 million a year ago
• Daily requests grew roughly 7x in a year
• The Gemini APIs now process about 19 billion tokens per minute
• Capital spending this year: $180-190 billion, up from $31 billion in 2022

That last one deserves a second look. Capex grew 6x in four years, much of it going to Google’s eighth-generation TPUs (the 8t and 8i). The bet is that inference demand keeps growing at this slope.

Gemini 3.5: usable on day one, not “coming soon”

The headline model release was Gemini 3.5 Flash, generally available on opening day across the Gemini app, AI Mode in Search, the API, the enterprise platform, and the Antigravity development environment. Google’s numbers: 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas, beating the older Gemini 3.1 Pro while running about 4x faster than other frontier models. A larger Gemini 3.5 Pro is in internal use and expected next month.

Flash first, Pro later. That’s the reverse of how these launches used to go. My read: when a model has to carry Search-scale traffic, inference cost matters more than leaderboard position. You ship the cheap fast version into every entry point first, and the flagship becomes the supplement.

There’s also an entirely new model family, Gemini Omni, built to generate output from any input, starting with video. Unlike Veo, which only takes text, Omni Flash accepts mixed text, photos, video, and audio as input, and Google made a point of its handling of physics (gravity, fluids). Generated videos carry the invisible SynthID watermark, and it’s free in YouTube Shorts Remix for users 18 and up.

Spark: Google’s always-on personal agent

Gemini Spark is the most ambitious thing announced: a personal agent that runs 24/7 on Google Cloud virtual machines, working even when your phone is off. It connects to Gmail, Docs, and Sheets, plus third-party services like Canva and Instacart. It can draft emails, build study guides, and watch your statements for hidden credit card fees.

Gizmodo called it Google’s answer to OpenClaw, and that framing is fair: the open-source community made always-on personal agents a thing, and now the big platforms are shipping walled-garden versions. Google’s rollout is cautious. Trusted testers first, then a beta for US AI Ultra subscribers the following week, and Spark is designed to check in before taking major actions.

A companion feature called Daily Brief works overnight, going through your inbox, calendar, and tasks to build a prioritized digest by morning. That one started rolling out to US subscribers on launch day.

Search: the biggest surgery in 25 years

Google’s own words: the biggest upgrade to the search box in over 25 years. The new box takes text, images, files, videos, and Chrome tabs at once, and AI Overviews and AI Mode are merging into a single experience. Two scale numbers: AI Overviews now reaches more than 2.5 billion monthly users (Pichai on CNBC), and AI Mode has passed 1 billion.

The bigger change is Search agents. Information agents, coming this summer to Pro and Ultra subscribers first, monitor a topic across the web around the clock and send back synthesized updates. Add generative UI, which builds custom result layouts on the fly and arrives free for everyone this summer, and the search results page as we know it has been redrawn.

Ynetnews pointed out the cost: this rework aims squarely at OpenAI and Perplexity, but it puts Google’s own ad revenue and publisher traffic at risk. And there’s a counter-signal worth noting: Gigazine reported that DuckDuckGo’s user base surged right after I/O. Some people are voting with their feet against AI search.

One practical thing for developers

Antigravity moved to 2.0, now a desktop application that orchestrates several agents in parallel, with a CLI and SDK. The part that matters: Google is transitioning the old Gemini CLI to the Antigravity CLI, and on June 18 the legacy tool stops serving requests for AI Pro, Ultra, and free Code Assist accounts. If your scripts or CI still call Gemini CLI, you have a month to migrate.

AI Studio got a big one too: build complete native Android apps from prompts, preview them in an embedded emulator, and eventually publish straight to the Play Store. The path from vibe coding to a shipped app is now a single tool.

What ships when

Product	Who gets it	When
Gemini 3.5 Flash	Everyone	Live (May 19)
Gemini 3.5 Pro	—	Next month
Omni Flash	AI Plus / Pro / Ultra	Live
Spark	US Ultra subscribers	Beta next week
Search agents	Pro / Ultra	This summer
Generative UI	Everyone (free)	This summer
Gemini CLI forced migration	Developers	June 18

Our take

What separates this I/O from the last two years is availability. Gemini 3.5 Flash and Omni Flash shipped the same day they were announced, not as recorded demos. Google’s play is clear by now: use its own distribution (Search, Android, Workspace, YouTube) to put agents into every entry point, so you’re using AI before you’ve decided whether you want to.

Two open questions, though. First, nobody has solved trust for always-on agents like Spark, and Google knows it, which is why the rollout starts with trusted testers. Second, the Search rework cuts into Google’s own ad lifeline, and publishers will lose more traffic to AI Overviews and Search agents. We run a content site, so we’re downstream of this one. We’ll be watching how it plays out over the next few months.