Google Rolls Out Gemini 3.5 Flash, Pushing AI Toward Autonomous “Worker” Models

Google introduced Gemini 3.5 Flash on Tuesday at its annual I/O developer event, positioning the release as a move away from chat-focused assistants toward AI that plans, builds and carries out complex tasks with little human direction. The company says the new model can orchestrate software builds, oversee extended research projects and — in internal trials — assemble an entire operating system. Google made Flash the default in its Gemini app and enabled it in Search’s AI Mode, signaling the company wants these kinds of autonomous agents to reach both developers and everyday users.

DeepMind’s chief technologist, Koray Kavukcuoglu, told reporters the model balances response quality with very low lag, and that it surpasses the prior frontier model, 3.1 Pro, on a range of technical tests including programming and multimodal reasoning. He described Flash as roughly four times quicker than previous frontier models, and said an optimized build runs at about a dozen times the speed while maintaining the same output quality. That responsiveness, the company says, is important when multiple autonomous assistants operate simultaneously on long-running jobs.

On the I/O stage, Google engineer Varun Mohan walked attendees through a demonstration in which autonomous agents divided work into pieces and later merged their outputs to produce a working operating system inside Antigravity, Google’s integrated environment for agent development. The company also unveiled Antigravity 2.0, a desktop application intended to let teams design and monitor agent-based workflows, a shift from single-chat interactions to an environment where software helpers can access tools and run processes over time.

Google framed Flash and the forthcoming 3.5 Pro as complementary: Pro is meant to do heavy-duty planning and deeper reasoning, while Flash functions as the fleet of hands-on subagents that execute tasks and call out to external tools. “Think of Pro as the strategist and Flash as the squad that carries out the orders,” said Tulsee Doshi, Google’s senior director and head of product, in an interview. She added that organizations will decide where they want the larger-model thinking versus which jobs are better handled by faster, tool-oriented agents.

Partners are already experimenting with the new model. Banks and financial-technology firms have run multi-week processes with reduced human oversight, and data teams are using the model to surface patterns in dense datasets, Google said. Flash can stay active for hours to pursue objectives autonomously, though Doshi noted it will pause and request human guidance at key decision points or when it encounters permissions questions that need judgment calls.

The expansion of agent-capable models has drawn scrutiny. Google is defending itself through stronger cybersecurity and hazardous-materials guardrails and says it has tuned the model to engage more carefully with sensitive queries rather than simply shutting down. The company also faces an active lawsuit tied to last year’s interactions between a user and a prior Gemini model that preceded a tragic outcome. Google said the new safeguards and deployment controls are part of its effort to reduce risk as these capabilities become more widely available.

Gemini 3.5 Flash is available now through Antigravity, the Gemini API and Gemini Enterprise, and is embedded in the Gemini mobile app and Search’s AI Mode. Google also plans to use Flash to power Gemini Spark, a round-the-clock personal assistant aimed at helping consumers manage email and other daily tasks.