Google AI Updates: May 20, 2026

1. Gemini 3.5 Flash ships with Antigravity-backed subagents and frontier-class scores

Google. Google released Gemini 3.5 Flash at I/O with Gemini 3.5 Pro slated for next month, framing Flash as the primary agent-focused model in the family. Reported benchmarks include 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, and 83.6% on MCP Atlas, with output token generation roughly four times faster than competing frontier models. The model is wired into Google Antigravity to deploy collaborative subagents for long-horizon coding, financial preparation, and legacy codebase transformation, and ships immediately into the Gemini app, AI Mode in Search, the Antigravity platform, the Gemini API, and enterprise surfaces. Source

2. Sundar frames the keynote as the “agentic Gemini era,” with Spark as the persistent agent layer

Google. Sundar Pichai’s keynote anchored I/O 2026 around Google’s move from text prediction to action, with monthly token volume cited as 7x year-over-year to more than 3.2 quadrillion and Gemini app monthly users roughly doubling from 400 million to 900 million. The headline agent product is Gemini Spark, a 24/7 personal AI agent that handles background tasks across Google surfaces, rolling out to trusted testers this week. Pichai also introduced Gemini Omni Flash for unified video, image, and text generation across the Gemini app, Google Flow, and YouTube Shorts, and previewed Antigravity 2.0 as a standalone platform for developing and managing autonomous agents. Source

3. AI Ultra splits into a $100 tier, $200 plan drops in price, and Spark gates on Ultra

Google. Google restructured its AI subscription stack at I/O, introducing a $100 AI Ultra plan aimed at developers, technical leads, and advanced creators with 5x higher usage limits than the Pro plan, Gemini 3.5 Flash, 20TB of storage, and YouTube Premium individual. The existing top-tier Ultra plan drops from $250 to $200 while keeping its 20x-over-Pro usage limits, and Gemini Spark rolls out as a beta to U.S. Ultra subscribers next week. Project Genie expands to all eligible $200 Ultra subscribers globally, AI Inbox drops down from Ultra-only to Plus and Pro, and a new Daily Brief agent ships morning summaries to subscribers across tiers. Source

4. Workspace gets voice-native Gmail, Docs, and Keep, plus a new Google Pics image tool

Google. Google brought voice-first interaction into Workspace at I/O, with Gmail Live enabling voice-activated inbox search (“What’s my flight’s gate number?”), Docs Live operating as a conversational drafting partner, and Keep transcribing voice notes into organized lists. The release also introduces Google Pics, a Trusted Tester image creation and editing tool with object segmentation, in-image text editing with translation, and real-time collaborative editing. AI Inbox in Gmail expands to AI Plus and Pro subscribers with personalized draft replies and inline access to relevant files, and Gemini Spark is integrated across Workspace apps to take action on the user’s behalf when permitted. Source

5. AI Mode in Search crosses one billion monthly users, with queries triple traditional length

Google. Google published one-year-in metrics for AI Mode in Search, reporting more than one billion monthly active users globally and AI Mode query volume more than doubling every quarter since launch. The average AI Mode search is roughly triple the length of a traditional Search query, more than one in six U.S. searches now uses voice or images, and image-based searches are growing 40% month over month. Planning queries are growing 80% faster than overall AI Mode usage and brainstorming searches 30% faster, with intent-style prefixes like “where to,” “where should I,” and “ideas for” leading the new behavior mix. Source

6. TPU 8t and 8i debut as the new training and inference silicon underneath the I/O announcements

Google. Pichai used the keynote to introduce the eighth generation of Google’s custom AI accelerators, splitting the family for the first time into TPU 8t for training and TPU 8i for inference, with Google citing roughly 2x better performance-per-watt versus the prior generation. The split mirrors the broader industry move toward dedicated inference silicon as agentic workloads push token volumes higher, and Google positioned the new TPUs as the substrate behind Gemini 3.5, Gemini Omni Flash, and the Spark agent fleet. The chips are already running internally and underpin Google’s claimed pricing advantage on Gemini 3.5 Flash, billed at “less than half the price” of competing frontier models. Source