Google I/O 2026 is May 19–20. Sixteen days away.
This year feels different from previous I/Os. The announced session topics — agentic coding, on-device AI inference, agent-native platforms — suggest Google is making a coordinated move to own the AI development stack end-to-end. Not just a model announcement or a framework update, but a claim on where AI-powered development happens.
Here's what's coming, broken down by what actually matters for developers.
TL;DR
- Gemini 4: 2M-token context window — entire large codebases fit in context without RAG
- Android 17 AI Core API: Third-party apps can call on-device Gemma Nano for free, no network, no API costs
- Google's agentic coding tool: Claude Code competitor built on Firebase Studio + Stitch + MCP
- Flutter GenUI: AI-generated adaptive interfaces, details at I/O
- Schedule: Main keynote 10am PT May 19, Developer keynote 1:30pm PT May 19
Schedule: What to Watch and When
| Date | Time (PT) | Content |
|---|---|---|
| May 12 | TBD | Android 17 user features preview |
| May 19 | 10:00 AM | Main Keynote — AI, Gemini model reveals |
| May 19 | 1:30 PM | Developer Keynote — Firebase, Android Studio, agentic coding |
| May 19–20 | All day | Sessions, codelabs, Q&A |
The developer keynote is the one to prioritize if you only have time for one. Google has officially named three developer stories for this I/O: Gemini 4, Android 17 AI integration, and the agentic coding tool.
Gemini 4: What a 2M-Token Context Window Actually Changes
The headline spec for Gemini 4 is a 2 million token context window — roughly 10x the current context limits of Claude Opus 4.7 (200K) and GPT-5.5 (256K).
Context window comparison (May 2026)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
GPT-5.5 256K tokens
Claude Opus 4.7 200K tokens
Gemini 4 (TBA) 2,000K tokens ← ~10x
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
In practical terms: at 2M tokens, an entire large codebase fits in context without retrieval-augmented generation. That's not a marginal improvement — it changes how you'd architect AI-assisted development workflows.
The current standard approach for large-codebase analysis looks like this: identify relevant files → chunk → embed → build a RAG pipeline → search → pass results to the model. That pipeline takes real engineering effort to build and maintain.
With 2M context, you hand over the repository directly. RAG still makes sense in specific scenarios — very large codebases, real-time retrieval requirements — but for mid-size projects and prototyping, it stops being a required step.
Gemini 4 will be previewed at I/O, with broader availability expected in late 2026 or early 2027. API access will roll out through Google AI Studio and Vertex AI.
Android 17 AI Core API: On-Device Inference Without the Tradeoffs
For Android developers specifically, the most immediately impactful announcement is the Android AI Core API in Android 17.
Until now, Android apps had two options for AI inference: call a cloud API (network dependency, per-token costs) or bundle a model locally (complex deployment, larger APK). Both have friction.
The Android AI Core API takes a different approach.
// Conceptual Android AI Core API usage
AIManager aiManager = AIManager.getInstance(context);
AIInferenceSession session = aiManager.createSession(
AIModel.GEMMA_NANO,
AISessionConfig.builder()
.setTask(AITask.TEXT_GENERATION)
.build()
);
String result = session.run("Summarize this user note: " + input);
Third-party apps can call the Gemma Nano model that Android manages at the OS level — no model bundling, no cloud calls. The system handles the model; developers call an API.
Three concrete implications:
No API costs. On-device inference means no per-token billing. For consumer apps where AI features need to scale to millions of users, this changes the cost model significantly.
No network latency. Works offline. Relevant for keyboard apps, real-time translation, accessibility tools, any feature where response time is user-visible.
No APK size increase. The OS manages the model file. Developers don't include it in the app bundle.
The Developer Preview for Android 17 will be available starting May 19, the same day as the keynote.
Google's Agentic Coding Tool: The Firebase Studio Play
The session descriptions for I/O 2026 name a Google agentic coding tool as one of the three main developer stories — positioned explicitly as a competitor to Claude Code and Cursor.
What's known about the stack:
- Firebase Studio: The hub for moving from AI prototyping to production deployment on Google Cloud
- Stitch: UI component generation and editing layer
- Antigravity: Full-stack application building tool within Firebase
- MCP integration: Model Context Protocol for connecting external tools
Firebase is described in session copy as evolving into an "agent-native platform," with an explicit path from AI Studio prototyping through to Google Cloud production. Android Studio sessions cover Gemini integration directly into the development workflow.
The differentiator over Claude Code and Cursor would likely be deep Google Cloud integration — if you're already running on GCP, Firebase, or using Google Workspace, the toolchain integration could be meaningfully tighter than what third-party tools offer. Whether the actual coding agent capability is competitive is something we'll know after the demos.
Flutter GenUI and Other Developer Updates
Flutter GenUI is a new Flutter capability being unveiled at I/O. The premise: instead of fixed UI components, the app dynamically generates and adapts its interface based on user context — input patterns, screen size, accessibility settings. Implementation details haven't been fully disclosed pre-I/O.
Flutter also gets general performance updates. For developers building cross-platform apps, this will be worth tracking on the second day of sessions.
What to Do Before May 19
1. Set up Android 17 Developer Preview environment If you want to start testing the AI Core API immediately after the keynote, have Android Studio (latest canary) and Android 17 preview SDK ready in advance.
2. Join Firebase Studio waitlist Access at firebase.google.com/studio. Being on the waitlist puts you closer to the front of post-I/O access.
3. Secure a Gemini API key When Gemini 4 previews drop, quota will be competitive. Having a Google AI Studio account set up avoids the setup delay when access opens.
4. Check the session catalog io.google/2026 has the session list. With two days of content across multiple tracks, knowing which sessions match your stack before the event makes the replay queue manageable.
The Bigger Picture
Gemini 4's 2M context, Android AI Core, and the Firebase agentic coding tool are each independently significant. Together they point in the same direction: Google is building toward owning the full AI development pipeline — from cloud infrastructure to on-device inference, from prototyping to production deployment.
Whether the I/O announcements deliver on that ambition is a question for May 19. But the session structure alone signals this isn't a scattered product update — it's a coordinated positioning move.
Worth watching closely.
Sources
- Get ready for Google I/O 2026 — Google Developers Blog
- Google I/O 2026 Developer Preview: Gemini 4, Android 17, Agentic Coding — Abhishek Gautam
- What to Expect from Google I/O 2026 — Gadget Hacks
- Gemma 4: The new standard for local agentic intelligence on Android — Android Developers Blog
- Google I/O 2026 AI Announcements: Gemini 4 & Beyond — mejba.me
Related reading:
- Claude Code April 2026 Update — Current state of the tool Google's agentic coder is competing with
- MCP Hits 97M Installs: The Linux Foundation Takeover — Background on the MCP standard Google's tool is adopting