Google announces agent-optimized Gemini 3.5.Flash and a do-anything model called Omni

At the recent I/O event, Google showcased the impressive evolution of its AI technology with the introduction of Gemini 3.5 Flash. Just a year ago, the conversation was centered on the 2.5 version, but significant advancements have since unfolded, culminating in this latest iteration. Gemini 3.5 Flash is now being integrated into a variety of Google products, with the company asserting that this model surpasses its predecessor, the Pro version, in performance. This release marks a pivotal moment for Google, as Gemini 3.5 Flash is touted to deliver cutting-edge intelligence while maintaining an efficiency that could make intricate agentic tasks feasible on a larger scale. Tulsee Doshi, the senior director of product management for Gemini, emphasized that the innovations within Gemini 3.5 Flash are embedded across multiple Google applications, indicating that this is merely the beginning of a larger transformation. The landscape of generative AI is challenging, often seen as a costly endeavor as companies strive for improved efficiency. This challenge intensifies when creating agentic experiences that require prolonged operations to accomplish complex tasks. With Gemini 3.5 Flash, Google appears to be making significant strides towards addressing these issues. The new model boasts an impressive output capacity of nearly 300 tokens per second, while its benchmark results align with those of larger models such as the 3.1 Pro, which operates at just a quarter of that speed. Doshi highlighted that substantial advancements in pre-training have been achieved with Gemini 3.5 Flash, largely informed by user feedback from developers utilizing Gemini models. "With our post-training approach, we are beginning to harness valuable insights from user interactions, exemplified by Antigravity," stated Doshi. "This is reflected in the improved code and tool performance. We anticipate that Gemini 3.5 Pro will further enhance these capabilities, and that future iterations will continue to bridge the performance gap with the Pro series."