Google Unveils Gemini 3.5 and New AI Agents at I/O 2026
CNBC reported Tuesday that Google unveiled a suite of new AI models and personal agent tools at its annual I/O developer conference, making its most assertive push yet to compete with OpenAI and Anthropic.
Gemini 3.5 Flash Takes Center Stage
The headline product is Gemini 3.5 Flash, a lighter, faster iteration of Google’s flagship model family. CEO Sundar Pichai described it as “remarkably fast” at a pre-event briefing with reporters. The company says the model delivers frontier-level capabilities at roughly half the cost of comparable offerings — and sometimes as low as one-third. Gemini 3.5 Flash will now serve as the default model inside the Gemini app and within Google’s AI-powered search mode globally. Google said the model carries stronger cybersecurity protections, making it less prone to generating unsafe outputs or wrongly refusing legitimate queries. A heavier variant, Gemini 3.5 Pro, remains in internal testing and is not expected to reach a broader audience until next month.
Spark Agent Targets Everyday Digital Tasks
On the agentic side, Google introduced Gemini Spark, a general-purpose AI agent built to reason across a user’s connected applications and execute tasks with minimal input. The company framed the product as a digital co-pilot capable of acting on a user’s behalf under their direction. Spark launches in beta next week, initially restricted to trusted testers and subscribers to Google’s AI Ultra tier. The move signals Google’s intent to deepen product integration across its platforms — something Wall Street has been pressing for as the company’s capital expenditure has climbed sharply.
Background: Pressure From Rivals Mounts
Google’s I/O announcements arrive at a moment of intensifying competitive pressure. OpenAI and Anthropic are both said to be preparing initial public offerings as soon as this year, drawing heavy investor attention and pushing valuations to record levels. Anthropic’s recently released Mythos model drew particular notice after reportedly uncovering thousands of previously undiscovered software vulnerabilities across global infrastructure. That development raised the stakes considerably for Google’s own model roadmap.
Omni World Model Expands Into Physical Simulation
Google also announced Omni, a world model designed to simulate physical environments and predict outcomes based on user inputs. The technology has long been a focus of Google’s DeepMind research division, where it has found applications in robotics and gaming. Omni will integrate across Gemini Flash, the Gemini app, Google Flow, and YouTube Shorts, supporting image and audio generation. Users will be able to direct Omni to alter video footage, insert new characters, or modify on-screen actions entirely through natural language prompts.
Read Next: OpenAI Targets $500B Valuation as IPO Speculation Grows
