Google AI Updates: April 4, 2026

1. Gemini API Adds Computer Use Tool Support in Preview Models

Google. The Gemini API now supports Computer Use tools in the gemini-3-pro-preview and gemini-3-flash-preview models, allowing developers to build agents that interact with desktop and browser interfaces programmatically. This positions Gemini alongside Anthropic’s Claude as a provider of computer-use capabilities via API. Source

2. Android AICore Developer Preview Brings Gemini Nano 4 On-Device

Google. Google launched the AICore Developer Preview for Android with two new on-device model variants: Gemini Nano 4 Fast (based on Gemma 4 E2B) and Gemini Nano 4 Full (based on Gemma 4 E4B). Google claims 4x faster inference and 60% less battery consumption compared to predecessors, with code written now forward-compatible with production Gemini Nano 4 devices later this year. Source

3. AI Edge Gallery and LiteRT-LM Launch for On-Device LLM Deployment

Google. Google released the AI Edge Gallery app for iOS and Android, enabling on-device AI experimentation, alongside LiteRT-LM, an open-source inference framework for deploying LLMs on edge hardware. The E2B model runs under 1.5 GB RAM and processes 4K tokens in under 3 seconds on a Raspberry Pi 5, making production-grade on-device inference accessible to mobile and IoT developers. Source

4. Agent Development Kit for Go Hits 1.0 with OpenTelemetry and Human-Approval Flows

Google. ADK Go 1.0.0 shipped with native OpenTelemetry integration for agent observability, self-healing logic via plugins, and a RequireConfirmation human-approval flow for sensitive tool operations. The Go SDK reaches parity with the Python ADK for building production agent systems on Google’s framework. Source

5. ADK SkillToolset Introduces Progressive Disclosure for Agent Context Management

Google. Google published the Developer’s Guide to Building ADK Agents with Skills, introducing the SkillToolset class with a three-level progressive disclosure architecture: L1 Metadata (~100 tokens), L2 Instructions (<5K tokens), and L3 Resources. The approach reduces baseline context token usage by approximately 90% compared to monolithic agent prompts, addressing a key scalability bottleneck for multi-tool agents. Source