Two IA Innovations from Google: Gemini 3.5 Flash Can Use Your PC, and the 'Circle to Search' Feature from Android Comes to Desktop

Google has introduced a new tool in Chrome that allows users to select a portion of the screen and ask Gemini for information about what is depicted. The feature, called "Select from Screen", comes along with a broader update that brings computer control capabilities directly into the Gemini 3.5 Flash model.

"Select from Screen" replicates the functioning of Circle to Search, which is already widespread on mobile, but applied to desktop navigation: you just need to select an area of the page displayed in Chrome to obtain explanations, translations, or insights about the highlighted content from Gemini, without copying text or changing screens. For those who browse a lot and often need to verify information on different pages, the main advantage lies precisely in its speed: there’s no need to open a new tab or manually paste any content into a separate chat.

The innovation is part of a more structural change announced by Google: Computer Use, the feature that enables a Gemini agent to interpret screenshots and perform actions like mouse clicks, scrolling, and keyboard typing, becomes an integrated tool in Gemini 3.5 Flash. Previously, it was available only through a separate model in preview.

With Computer Use, the application developed must carry out the actions suggested by the model, capture the resulting screen, and send it back to Gemini, creating a continuous loop until the requested task is completed. Google indicates potential uses such as repetitive form filling, testing applications, conducting research across multiple sites simultaneously, and more extensive business workflows.

Gemini 3.5 Flash, announced in May as Google's quick model for programming and multi-step workflows based on agents, supports a context window of up to one million tokens and a maximum output of 65,000 tokens, with adjustable reasoning levels to balance response quality, latency, and costs.

For computer control scenarios, the model has undergone specific training against manipulation attempts. Google has implemented security measures that require user confirmation before sensitive or irreversible actions, as well as an automatic workflow interruption when a suspicious prompt injection attempt is detected, with configurable protections for areas such as financial transactions or the modification of sensitive data.

Google is not the first to bring computer control to its platform: Anthropic already offers it through Claude, while OpenAI continues to refine the performance of its models in this area. Microsoft has also applied the concept to business workflows, integrating it into the Researcher agent of Microsoft 365 Copilot.