Google I/O '26 Keynote

Tech/Product concepts highlighted

• “Full stack, AI-first” strategy (Google): AI innovation spanning custom silicon + secure foundations, research/model development, and AI features across products/platforms.

Gemini adoption & scale metrics

• Tokens processed grew from 9.7T/month (two years ago) → 480T/month (last I/O) → 3.2 quadrillion tokens/month (today).
• Over 8.5M developers building with models monthly.
• 13 Google products now have 1B+ users; 5 exceed 3B users.

AI Search upgrades

• AI Overviews: 2.5B monthly users.
• AI Mode: exceeded 1B monthly users; queries doubling every quarter since launch.
• Major UX changes:
  • A new intelligent search box with multimodal inputs (text/images/files/videos).
  • A more seamless “AI Search” experience that continues the conversation using AI Overviews + AI Mode together.
• Search agents: run multiple background information agents 24/7 in Search (e.g., finance tracking, apartment hunting, sneaker drop monitoring).
• Agentic coding in Search: generate custom interactive “generative UI” by planning/constructing UI + executing code in a secure harness (built on Antigravity concepts).
• Stateful “mini-app” experiences: ongoing trackers/dashboards (example: weekend planner with calendar/weather/driving time integrations).

Gemini app growth & creative capabilities

• Gemini app MAU grew from 400M (last I/O) → 900M (today); daily requests up 7×.
• Personal intelligence: more customized responses via secure connections to Google apps (Gmail, Photos, etc.).
• Image generation: 50B+ images generated with Nano/“banana” models.
• Gemini Omni (world model/editor) introduced:
  • A new model family concept: create/edit realistic videos/images/interactive simulations from multimodal inputs.
  • Focus on world understanding (e.g., simulating physics like gravity/kinetic energy) and iterative editing with conversational guidance.
  • Gemini Omni Flash launched as the first Omni-family model; available across products.
• Gemini 3.5 Flash / 3.5 Pro direction:
  • 3.5 Flash: positioned as faster, strong on coding, and agentic long-horizon tasks.
  • Claims: ~4× faster than other frontier models, plus Antigravity optimization improvements (see Antigravity below).

Docs via voice (“Docs live”)

• In the Gemini app: create/edit docs by speaking freely (no precise prompts required).
• Demo highlights:
  • Pull context from Drive and email subjects.
  • Generate tables/formatting and bold story notes.
• Rollout:
  • To Pro/Ultra subscribers this summer.
  • Voice features extend to Gmail + Google Keep.

Content authenticity / transparency tools

• SynthID watermarking (invisible):
  • Already watermarked 100B+ images/videos and 60k years of audio assets.
• SynthID detector used in Gemini app.
• Content credentials verification:
  • Indicates whether content origin was AI vs camera and whether edited with genAI tools.
• Expansion:
  • Verification tools coming to Search and Chrome (e.g., “circle to search” / right-click “was this generated with AI”).
• More partners adopting SynthID (examples mentioned): NVIDIA, OpenAI, Kakao, Levin Labs.

Antigravity: agentic development platform

• Antigravity 2.0: “agent-first” desktop app + harness for orchestrating multi-agent systems.

New developer interfaces

• Full CLI
• Native voice support with Gemini audio models
• Integrations with services like Android, Firebase, Google AI Studio

Core primitives

• Subagents
• Hooks
• Asynchronous task management

Performance/cost example

• Agents built a working OS from scratch in 12 hours using:
  • 93 subagents
  • 15k+ model requests
  • 2.6B tokens
  • Cost: < $1,000 in API credits
• Demonstrated OS running utilities and (after fixes) a DOOM-like run.

Scaling claim

• Tasks that take multi-day engineering efforts reduced to hours/minutes.

Speed/cost positioning

• Flash optimized to be ~12× faster in Antigravity (beyond the raw model’s headline speed).

Gemini Spark: personal always-on AI agent

• A dedicated agent in the Gemini app that runs 24/7 on Google Cloud for long-running background tasks.
• Powered by Gemini 3.5 models + Antigravity harness.
• Works with tools under user control; emphasizes safety/secure rollout:
  • Beta rollout to trusted testers first
  • Then Ultra subscribers next week (US)

Demo examples

• Draft email by collecting info from Docs/Email/Chats (“ghostwriter” skill).
• Build and maintain a block-party RSVP tracker in Google Sheets, update via Gmail signals, follow up with reminders.
• Generate a Google Slides hype deck.
• Task creation via voice: Spark breaks brain-dump tasks into subtasks and runs concurrently.

Platform expansion plans

• Later:
  • Spark in Chrome as an agentic browser
  • A phone “agent home base” on Android

Agentic e-commerce (agent commerce)

• Universal Commerce Protocol (UCP): like HTTP for agent commerce—an open standard for shopping journey interoperability.
• Agent Payments Protocol (AP2): payment boundaries + accountability:
  • Guardrails (brand/product + spend limits)
  • Privacy-preserving, verifiable linkage + tamper-proof digital “paper trail”
• Universal Cart:
  • Works across Search, Gemini app, YouTube, Gmail, etc.
  • Background deal detection (price drop insights, stock alerts).
  • Reasoning example: compatible PC parts suggestions (e.g., motherboard socket mismatch).
• Rollout:
  • Universal Cart starting in US across Search + Gemini app this summer.
  • YouTube + Gmail later.

Creative tools + UI creation

• Picks: Workspace image creation/editing tool with object-aware editing (remove/resize elements), text editing/translation, and SynthID watermarking.
• Stitch upgrades:
  • Faster UI design from prompts
  • Real-time collaboration with voice/prompt refinements
  • Export to code/launch site
• Google Flow / Flow Music:
  • Multi-step creative agents (multiple actions at once)
  • “Vibe coding” for creative tools

XR / intelligent eyewear

• Android XR platform with Samsung + Gemini integration.
• Display glasses:
  • Hands-free overlays (info in-lens)
  • Developer ecosystem building
• Audio glasses launching this fall:
  • Gemini spoken into ear privately (no constant display needed)
  • Demo: navigation + context-aware ordering via DoorDash, and calendar updates based on missed messages

---

Reviews/Guides/Tutorials mentioned

• “Docs live” demo:
  • Functions like a guided tutorial: verbally request doc creation, pull context from Drive/email, then format into tables and refine narrative.
• Antigravity OS demo:
  • Effectively a developer tutorial/showcase: agents plan → generate code → run tests → iterate (then apply fixes for drivers to run an app).
• Search demo (generative UI):
  • Acts as an interactive guide: user asks a question, Search builds interactive visuals, and allows parameter exploration.

---

Main speakers / sources (as stated in subtitles)

• Sundar Pichai
• Matthias (mentioned as an AI-assistant author; with brother Tommaso)
• Demis Hassabis (DeepMind)
• Varun Mohan
• Liz Reid
• Josh Woodward
• Robbie Stein
• Vidhya Srinivasan
• Shiram (Shahr ab likely; presented XR segment host) / Shahr am Izadi (XR segment)
• Jay Kim (Samsung)
• Dennis (Gemini app segment; likely “Dennis” as referenced in subtitles)
• Holly (small business story: restaurant + Gemini/Antigravity hiring platform)
• Suz Chambers
• Josh Woodward (repeated earlier)