Google I/O 2026: Gemini 3.5 Flash Lands Today, a Personal AI Agent Called Spark, and a $5B TPU Cloud With Blackstone

Google used its I/O 2026 keynote in Mountain View today to fire a sustained barrage of AI launches at the gap that had opened up between Gemini and rival frontier systems from OpenAI and Anthropic. Sundar Pichai opened by calling the past year one of "hyper progress" and described Google Search as the company's "ultimate moonshot," then handed the stage to a cadence of demos that touched the model, the agent layer, the operating system, and the data center underneath.

Gemini 3.5 Flash is the immediate centerpiece. It rolls out today across the Gemini app, Search, the new Antigravity 2.0 developer suite, and the Gemini API, with Gemini 3.5 Pro following in June. Pichai showcased a steep performance jump: 76.2% on Terminal-Bench 2.1 (versus 70.3% for the previous Gemini 3.1 Pro), 83.6% on MCP Atlas for scaled tool use (up from 78.2%), and a leap to 1,656 Elo on GDPval-AA — the real-world agentic benchmark where 3.1 Pro had scored 1,314 and trailed Anthropic's Claude Sonnet 4.6. The model also clocks 289 tokens per second, which Google says is roughly four times the throughput of competing frontier systems.

The big new product on top of that model is Gemini Spark, a 24/7 personal AI agent that runs on private Google Cloud infrastructure even while the user is offline. Spark coordinates across Sheets, Gmail, Calendar, and Chrome, and is positioned for jobs like party planning, vendor management, long-running research, and shifting drafts as new information arrives. It launches in beta next week, gated initially to Google AI Ultra subscribers in the US, with an agentic browsing path through Chrome arriving later in 2026. Google sharpened pricing for the moment, adding a new $100 mid-tier and cutting the top-end Ultra plan from $250 to $200 a month.

The agent push is mirrored at the operating-system layer. Android Halo, shipping in summer 2026, gives Spark and third-party agents a dedicated home on the phone — a system-level surface for long-running tasks rather than a chat window crammed into the launcher. Pichai also showed off Universal Cart, a Gemini-powered shopping cart that follows users across Google surfaces, tracks sales and stock changes, and flags product incompatibilities; it launches in the US this summer. Ask YouTube, a conversational search experience for video, rolls out in the US during 2026.

Productivity got a voice-first overhaul. Docs Live and Gmail Live let Pro and Ultra subscribers compose, edit, and reply by talking out loud rather than typing, with the model handling structure, citations, and tone matching. Google Keep picks up the same voice creation surface, and Google Pics got a brief mention as a new Gemini-powered photos experience. Developers got their own headliner with Antigravity 2.0, an "agent-first" coding platform that Google says was used internally to build a working operating system from scratch for under $1,000 in token spend. It is available globally today.

Video and multimodal generation arrived through Gemini Omni and its faster sibling Omni Flash, the production version of the model that leaked the week before the keynote. Demis Hassabis framed Omni as Gemini's reasoning powers fused with a real grasp of physics: it can simulate kinetic energy, gravity, and other concepts, and turn scientific ideas into short explainer videos. "Down the road, Omni will be able to generate any output from any input," Hassabis said, repeating his frequent line that "artificial general intelligence is just a few years away." Omni Flash arrives across Google products this summer.

On the hardware side, Google moved more cautiously than rumors had suggested. The company previewed Android XR audio glasses launching in fall 2026 — notably with iPhone support — and built out the GoogleBook PC initiative, an AI-first laptop category developed jointly with Dell, HP, Lenovo, Asus, Qualcomm, and Intel, due this fall. Android 17 was treated as connective tissue rather than a tentpole: Gemini Intelligence is now baked into the system shell, with new digital-wellbeing controls, expanded AirDrop interoperability, and QR-based file sharing between Android and iPhone.

Two infrastructure items closed out the keynote and may matter more in the long run than any single product reveal. SynthID, Google's invisible watermarking system for AI-generated content, has been adopted by Nvidia, OpenAI, Kakao, and ElevenLabs as a cross-industry detection standard with content credentials baked in. And Google announced a $5 billion AI cloud venture with Blackstone, set to launch in 2027 with 500 megawatts of TPU capacity dedicated to external customers. Pichai closed by noting the Gemini app now has more than 900 million monthly active users — double a year ago — and is serving 9.7 trillion tokens per month, a scale that frames every reveal above as deployment, not preview.