Google I/O 2026 Preview: Gemini 3.2 Flash, Android 17, Gemma 4 — What Developers Get

Google I/O 2026 runs May 19-20 at Shoreline Amphitheatre in Mountain View, with a developer keynote on May 19 and technical sessions continuing through May 20. The conference is 6 days out as of today. Based on the Android Show pre-brief on May 12 and confirmed announcements from Google's pre-I/O communications, here is what developers can expect — and what it means for applications built on Google's AI and platform stack.

The headline story at I/O 2026 is not a single dramatic model release. It is deployment density: Gemini 3.2 Flash being rolled into Search, Maps, YouTube, Docs, Gmail, and Chrome for billions of users simultaneously. Capability at the frontier matters; distribution at Google's scale is a different kind of advantage.

Gemini 3.2 Flash: The Infrastructure Deployment Story

Gemini 3.2 Flash is Google's efficiency-optimised frontier model — the version designed for low latency and high throughput at Google's deployment scale rather than for maximum benchmark performance. Gemini 3.2 Ultra (the highest-capability variant) handles premium use cases; Flash handles the volume.

The confirmed I/O deployment scope: Gemini 3.2 Flash is replacing the previous AI layer in Google Search AI Overviews, Google Maps local information summaries, YouTube chapter and summary generation, Google Docs smart compose and contextual suggestions, and Gmail smart reply. Collectively, these surfaces serve several billion active users.

For developers, the Gemini 3.2 Flash API is the version you should default to for production applications where cost and latency matter. Based on the Gemini 2.5 Flash pricing as a reference point, Flash models run at approximately 1/8th the cost of Ultra per token. For applications that do not require graduate-level reasoning (most applications), Flash is the correct tier.

New in 3.2 Flash vs. 2.5 Flash: improved function calling reliability (the specific failure mode that makes agentic applications unreliable — incorrect JSON schema adherence in multi-turn conversations), better handling of long-context code analysis, and faster response times for structured output generation.

Android 17: Developer-Relevant Features

The Android Show on May 12 previewed Android 17's developer-facing changes:

On-device AI APIs: Android 17 formalises the on-device inference APIs that have been in developer preview since Android 15. The MediaPipe Tasks SDK is being deprecated in favour of a unified Android AI Core framework that handles model quantisation, hardware acceleration (NPU routing on Tensor chips), and memory management automatically.

Privacy Sandbox for AI: Federated learning improvements for on-device personalisation without data leaving the device. Relevant for developers building personalised features in regulated industries (healthcare, finance) where sending user data to cloud inference endpoints creates compliance complexity.

Edge-to-Cloud routing: New API for automatically routing inference requests between on-device models and Google Cloud Gemini endpoints based on model complexity, latency budget, and connectivity status. The developer writes one inference call; the framework decides where to run it.

Predictive back gestures and AI navigation: Minor UX changes, more significant for consumer apps than developer infrastructure.

Android 17 developer preview is expected to be released at I/O, with the stable release in Q3 2026.

Gemma 4: Open-Weights Model for Developers

Google is releasing Gemma 4, the fourth generation of its open-weights model series designed for on-device and self-hosted deployment. The Gemma series is Google's answer to Meta's Llama models — open weights, commercially usable, fine-tunable.

Expected improvements in Gemma 4 over Gemma 3: better instruction following, improved code generation, and a new 27B parameter variant that fits in GPU memory more efficiently through a quantisation scheme designed for 4-bit inference on consumer-grade hardware.

Gemma 4 models will be available on Google AI Studio, Hugging Face, and Kaggle. They are free to use and fine-tune for commercial applications under the Gemma terms of use.

For developers choosing between Gemma 4 and other open-weights options (Llama 3.1, DeepSeek V4, GLM-5.1): Gemma's advantage is the Keras and TensorFlow native integration, Google's support for TPU v7 fine-tuning, and the Vertex AI deployment pipeline that makes going from fine-tuned model to production endpoint straightforward for teams already in the Google Cloud ecosystem.

Android XR and the Glasses Developer Preview

Google is announcing a developer preview for Android XR glasses — the consumer hardware successor to the discontinued Google Glass Enterprise Edition. The developer preview is expected to be a limited hardware program with an SDK for building glasses-native applications.

The form factor: conventional eyeglass frames with a transparent display for AI-assisted overlay in the right lens, speaker and microphone for voice interaction, and connection to a paired Android phone or Pixel 10 Pro for compute offload.

The expected developer tools: Android XR SDK for building AI overlay applications, integration with Gemini on-device and cloud inference for real-time contextual information, and APIs for spatial audio and voice command handling.

The consumer launch timeline for Android XR glasses has not been confirmed. The developer preview is an SDK and limited hardware program for third-party developers to begin building before consumer availability.

Firebase AI and Cloud Tools Updates

Firebase is announcing several AI-related updates that are directly relevant to developers building production applications on Google's mobile and web platform:

Firebase AI Logic (formerly Vertex AI in Firebase): GA release of the Firebase AI Logic SDK, which provides direct access to Gemini models from mobile and web clients with built-in security rules, usage monitoring, and automatic API key management. The pitch: deploy Gemini in your app with the same security model as Firestore, without managing API keys in client code.

Firebase Genkit 2.0: GA release of Genkit 2.0, Google's TypeScript/JavaScript AI application framework. Genkit provides flow orchestration, tool calling, multi-model routing, and local development with Firebase emulator support. Genkit 2.0 adds streaming support, improved observability (traces integrated with Cloud Trace), and native MCP server integration.

Cloud Run AI inference endpoints: New Cloud Run configuration for AI inference workloads — automatic GPU attachment, model weight caching between cold starts, and request batching. Designed to reduce the operational complexity of running self-hosted models on Google Cloud.

What the Government AI Pre-Evaluation Agreement Means

Separately, on May 5, Google — along with Microsoft, xAI, OpenAI, and Anthropic — agreed to allow the US Commerce Department's Center for AI Standards and Innovation to evaluate new models before public release. This is a voluntary pre-release evaluation program, not a regulatory approval process.

For developers, the practical effect is minimal in the short term. The agreement creates a reporting and evaluation mechanism that may eventually influence which model capabilities trigger additional disclosure requirements. It is the beginning of a US AI governance framework, not a current operational constraint.

Key Takeaways

• Google I/O 2026: May 19-20: Developer keynote May 19; Gemini 3.2 Flash, Android 17, Gemma 4, Android XR developer preview, Firebase AI Logic GA — confirmed
• Gemini 3.2 Flash deployed at Google scale: Search AI Overviews, Maps, YouTube, Docs, Gmail, Chrome — billions of users; Flash API is the default tier for production developer applications (1/8th Ultra cost)
• Android 17: Unified Android AI Core framework replaces MediaPipe Tasks; Edge-to-Cloud inference routing API; Privacy Sandbox federated learning; developer preview at I/O, stable Q3 2026
• Gemma 4: Open-weights, commercially usable; new 27B variant with efficient 4-bit quantisation; Keras/TensorFlow/Vertex AI native; available on Hugging Face and Kaggle
• Firebase AI Logic GA: Gemini in mobile/web apps with Firestore-equivalent security rules; Firebase Genkit 2.0 with MCP server integration and Cloud Trace observability
• Android XR developer preview: SDK for AI overlay glasses applications; hardware limited program; consumer launch timeline not confirmed

For the competing Anthropic developer tools released the same week, read Anthropic Leases SpaceX Colossus 1: 220K GPUs, Claude Rate Limits Doubled. To compare current AI model APIs and pricing, use the LLM API Pricing Tracker.