Google I/O 2026: Android 17, Gemini Agentic Coding Preview

Google I/O 2026 is scheduled for May 19–20 at Shoreline Amphitheatre in Mountain View. Developer registration opened in late March. The keynote runs on May 19, with technical sessions and codelabs continuing through May 20.

This is the context: Gemini 3.1 Ultra shipped before I/O (April 2026), which means Google is using I/O not to announce the model itself but to announce what developers can build with it. That is a different session structure than prior years. When the flagship model drops 5–6 weeks before I/O, I/O becomes the developer tooling showcase, not the model announcement event.

What Google Has Confirmed for I/O 2026

Google's official I/O website and session agenda (released April 8) confirms four main tracks:

Android 17: Google confirmed Android 17 developer preview 3 ships at I/O. The headline features confirmed in earlier developer previews: predictive back gesture refinement, Compose UI updates, health and fitness APIs, and the AI Core on-device model runtime. The AI Core integration is the structural change — it gives Android apps a standardised interface to run on-device AI models without bundling the model binary in the APK. Third-party models (including smaller versions of Gemini) can be loaded via AI Core.

Gemini for developers: Google has confirmed a full Gemini developer track at I/O. Based on the session titles published (topics without speaker details), this includes: Gemini API updates, AI Studio new features, Vertex AI agent builder announcements, and a session specifically titled "Building production-ready agentic workflows with Gemini." The agentic workflow session is the one to watch — it likely previews the Google equivalent of Anthropic's Claude Cowork and OpenAI's Operator.

Flutter and Dart updates: Flutter 4.0 is expected at I/O. The headline Flutter change that has been previewed: AI-assisted widget generation (Gemini writes Flutter components based on design descriptions), significantly improved web performance via a Wasm-compiled renderer, and multi-window support for desktop.

Web platform: Google has a track on web platform advances, expected to cover updates to Chrome's Prompt API (the browser-native LLM access API in Chrome 136+) and the progression of WebGPU from experimental to stable in Chrome.

The Agentic Coding Focus: What It Means

Prior Google I/Os were structured around Firebase, Flutter, and Android as the three developer pillars. I/O 2026 has shifted. The session agenda shows more AI developer tool sessions than any prior I/O. The structural theme: Google is repositioning itself as an AI-native developer platform, not a Firebase-plus-Android platform that also does AI.

The specific sessions that signal this shift:

"Gemini in Android Studio": Google's Android IDE now has Gemini integration for code completion, test generation, and — new for I/O — a full agent mode that can read your project, understand existing architecture, and propose multi-file refactors. This is direct competition with Cursor and GitHub Copilot's workspace feature.

"Project Astra in developer preview": Project Astra is Google DeepMind's multimodal agent that can see your screen, hear your voice, and take actions. At I/O 2025, it was a research demo. At I/O 2026, the session is in the developer track — suggesting an API or SDK release.

"Gemini API long-context workflows": A technical session on using the 2M token context window for production workloads. Based on what developers need to know: how to structure prompts efficiently in long context, how to use the context cache feature (which avoids re-processing the static part of a long prompt), and how to measure coherence degradation across the window.

Android 17: The AI Core Runtime Matters Most

Android 17 has been in developer preview since January. The AI Core runtime is the feature that matters most for developers building AI-adjacent apps.

What AI Core does: Provides a standardised on-device API for accessing lightweight AI models — text classification, entity extraction, image segmentation, and small language model tasks — without internet connectivity. Apps call AI Core's API; AI Core handles model loading, inference, and memory management. Apps do not need to know which model is running.

Why this matters: Currently, every app that wants on-device AI bundles its own model (TensorFlow Lite, ONNX, or MediaPipe). This bloats APK sizes and creates a fragmented model management problem. AI Core centralises this. Google, Samsung, and device OEMs can preinstall optimised models for their hardware. App developers call the API.

The Gemini Nano connection: Gemini Nano (Google's smallest model, optimised for mobile) runs through AI Core on Pixel 9 and some Samsung Galaxy S25 devices today. AI Core at Android 17 extends this to the wider Android ecosystem, with a published API surface that non-Google apps can use. If I/O confirms full third-party access to AI Core's Gemini Nano tier, it gives Android developers an on-device LLM that they do not have to ship, maintain, or pay API costs for.

What to Expect Announced but Not Yet Leaked

Based on Google's pattern from prior I/Os and the session themes published:

NotebookLM Pro: NotebookLM (the long-document Q&A tool) has been a consumer product. Expect a Pro tier with API access announced at I/O — this is the enterprise document analysis product that competes with Claude Cowork on document-heavy workflows.

Workspace AI expansion: Google Docs, Sheets, and Slides have Gemini integration. Expect announcements about extending this to Google Drive bulk processing (summarise 500 documents in Drive) and Google Meet real-time agent capabilities (live meeting assistant that can answer questions from company knowledge bases during the meeting).

Gemini API context caching pricing: The context caching feature (which lets you pay once to cache the static part of a long context) was announced in beta. Expect pricing and general availability announcement at I/O. This is a major cost reduction for developers using the 2M context window repeatedly on the same document.

Chrome Prompt API stable: Chrome's Prompt API gives websites direct access to an on-device language model without API calls. It has been in origin trial in Chrome 136+. An I/O announcement about stable availability would mean any website can run a basic LLM in the user's browser without server infrastructure.

Session Schedule and How to Watch

Google I/O 2026 keynote: May 19, 10:00 AM Pacific (17:00 UTC). Livestreamed on YouTube and io.google.

Developer sessions: May 19–20, sessions run in multiple rooms simultaneously. The full session catalog with codelabs is available on the I/O website. For developers who cannot attend in person, all sessions are recorded and released on YouTube the same day.

The sessions worth blocking your calendar for (based on session titles in the published agenda):

1. Opening keynote (May 19, 10:00 AM PT) — model and product announcements
2. Developer keynote (May 19, ~1:30 PM PT) — the actual developer tools announcements
3. Building production-ready agentic workflows with Gemini — the agentic coding session
4. Android 17 and AI Core deep dive
5. Gemini API long-context workflows

If you are evaluating Gemini for production use, the developer keynote and the long-context session are higher priority than the main keynote for actionable information.

Why This I/O Matters More Than Prior Ones

Google I/O 2024 and 2025 were Google playing catch-up. Google I/O 2026 is different: Gemini 3.1 Ultra leads on context window by a 10x margin, Android 17's AI Core is infrastructure-level (not a feature), and Project Astra moving from research demo to developer preview suggests Google is competing for the agentic developer market, not just the enterprise search market.

For developers deciding between Google Cloud and AWS or Azure for AI workloads, I/O 2026 is a real decision point. The Vertex AI announcements, the Gemini API pricing for context caching, and the agentic workflow tooling will determine whether Google Cloud is the better home for Gemini-native applications — or whether the tooling maturity gap vs OpenAI's and Anthropic's developer ecosystems persists.

Key Takeaways

• Google I/O 2026 is May 19–20 at Shoreline Amphitheatre — keynote May 19 at 10 AM PT, developer sessions run through May 20, all sessions livestreamed and recorded on YouTube
• Android 17 at I/O: AI Core runtime is the structural change — standardised on-device LLM API for Android apps, Gemini Nano access without bundling model binaries
• Gemini agentic developer tools: Gemini in Android Studio with full agent mode, Project Astra developer preview, production agentic workflow sessions — Google repositioning as AI-native developer platform
• Expected announcements: NotebookLM Pro API, Workspace Drive bulk processing, Gemini context caching general availability, Chrome Prompt API stable release
• Context: Gemini 3.1 Ultra shipped before I/O — I/O is the developer tooling showcase, not the model reveal; watch the developer keynote, not just the main keynote
• Decision signal: If Vertex AI agentic tooling announcements close the maturity gap with Anthropic's MCP ecosystem and OpenAI's Operator, Google Cloud becomes a more compelling AI infrastructure choice

Track AI API costs as new models launch with LLM API Pricing. Compare Gemini's current capabilities with Claude vs ChatGPT. For context on how AI chip supply enables the models being announced at I/O, read TSMC Q1 2026 and AI accelerator demand.