[I/O 2026] Gemini 3.5 ushers in the era of agentic models at Google

Google launches Gemini 3.5 Flash, an agentic AI model outperforming Gemini 3.1 Pro on coding and reasoning benchmarks, now available on Google AI Studio.

With Gemini 3.5, Google unveils its first family of models natively designed for agentic AI: systems that reason, plan, and act autonomously on long and complex tasks, going beyond simple responses.

The series kicks off with Gemini 3.5 Flash, presented as the most powerful model in the Flash lineage to date. It surpasses the previous Gemini 3.1 Pro on the most demanding coding and agentic tasks: Terminal-Bench 2.1 → 76.2% (vs. 70.3% for 3.1 Pro) MCP Atlas (multi-step workflows with tools) → 83.6% (vs. 78.2%) CharXiv Reasoning (advanced multimodal understanding) → 84.2%

Google also claims a four times higher throughput than other frontier models in output tokens per second, often at half the cost.

The model immediately becomes the default version of the Gemini app and Google Search's AI mode globally, and powers Gemini Spark, the personal agent that runs continuously on Google Cloud virtual machines. Developers and businesses access it via Antigravity, the Gemini API in AI Studio, Android Studio, and Gemini Enterprise. The 3.5 Pro version, already used internally, is expected next month. The entire family has been trained and evaluated according to the in-house Frontier Safety Framework, with strengthened cybersecurity protections and CBRN risks (chemical, biological, radiological, and nuclear), a reduction of abusive refusals on legitimate requests, and interpretability tools capable of inspecting internal reasoning before sensitive responses.

The Gemini experience is simultaneously completely redesigned through a new visual language dubbed Neural Expressive: fluid animations, vibrant colors, novel typography, and haptic feedback. The Gemini Live conversational mode is now integrated directly into Gemini, allowing users to switch from a typed question to a continuous spoken exchange without interruption, with the microphone having been revised to unfold a complex idea at its own pace without being cut off. Regional dialects and a choice of voices are planned soon. Finally, responses are formatted in real-time with images, interactive timelines, commented videos, and dynamic graphics rather than a block of text.

Neural Expressive is rolling out today on Web, Android, and iOS.