Google rolls out Gemini 3.5 Flash, an efficient model designed for agent tasks, and introduces Gemini Spark, a dedicated agent running in Google's cloud.

Gemini 3.5 Flash is rolling out across Google products starting today. The model can output nearly 300 tokens per second while achieving benchmark scores similar to larger frontier models like Gemini 3.1 Pro that output at a quarter of that speed. On code generation tests (Terminal Bench and SWE-Bench Pro) and UI control tasks (OSWorld-Verified), it substantially outperforms older Flash models and shows comparable or slightly better performance than Gemini 3.1 Pro and OpenAI's GPT 5.5.

Gemini Spark is Google's first dedicated agent, running 24/7 in Google's cloud using Gemini 3.5 Flash. It can execute tasks over time by pulling context from Drive, Gmail, and other services—such as tracking emails for digests, monitoring meetings for summaries and action items, or managing other recurring workflows. Google says it is designed to ask for approval before undertaking high-stakes actions.

The efficiency improvements in Gemini 3.5 Flash enable workflows previously too expensive to scale. Google's Antigravity IDE is being upgraded to version 2.0 with support for multiple parallel sub-agents spawned by Gemini 3.5 Flash. Gemini 3.5 Pro is in internal testing and should be ready for release next month. The model is coming to the Gemini app, the API, AI Studio, Android Studio, and Google's enterprise products.