Google Adds Native Computer Use to Gemini 3.5 Flash

According to the Google DeepMind blog post published Jun 24, 2026, computer use is now a built-in tool in Gemini 3.5 Flash, enabling the model to perceive and interact with screen content and GUIs so agents can operate across browser, mobile, and desktop environments. The blog states developers can build agents that use screenshots and visual understanding to navigate websites, click buttons, fill forms, operate enterprise software, and carry out multi-step workflows. The DeepMind model pages and accompanying evaluation PDF provide benchmark results and methodology for gemini-3.5-flash, including an OSWorld-Verified UI Control score of 78.4% reported on DeepMind's public model page and agentic benchmark results summarized in the model evaluation PDF.

Per the DeepMind evaluation PDF and the Google DeepMind blog, the computer use capability is integrated natively into the main Flash model rather than provided as a separate add-on. The evaluation methodology document describes benchmark suites used for agentic and UI tasks (Terminal-Bench 2.1, MCP Atlas, Toolathlon, OSWorld-Verified), and notes self-computed runs averaged over multiple trials for Gemini models. The PDF and model pages list harness and tooling details used for UI actuation (for example, pyautogui for actuation and the OSWorld docker and default 1080p resolution for UI control tests). The Google blog describes two optional enterprise safeguards: one that requires explicit user confirmation for sensitive or irreversible actions, and one that can automatically stop tasks if an indirect prompt injection is identified.

Companies adding visual screen-control primitives to language agents remove a major integration bottleneck that traditionally required per-application APIs or bespoke connectors. Industry-pattern observations: agents that operate via visual UI control typically combine robust visual grounding, stateful workflow planning, and reliable actuation libraries; they also need sandboxing and layered safety controls to limit unintended side effects. For practitioners, the availability of a native computer-use tool inside a frontier model like gemini-3.5-flash lowers engineering overhead for building automation across legacy GUIs but raises reproducibility and monitoring demands because visual actuation is more sensitive to UI changes and timing than API calls.

DeepMind and Google Cloud framing around an "agentic enterprise" places this capability alongside other Agent Platform and managed-agent features announced at Google I/O and Google Cloud events. Public benchmark numbers on the DeepMind model page and in the evaluation PDF position Gemini 3.5 Flash as a top performer on several agentic and coding metrics (the model page lists comparative rows for Terminal-Bench, MCP Atlas, and other suites). Observed patterns in similar releases: when vendors add native UI-control capability, customers prioritize auditability, human approvals, and access controls, and third-party tooling for replay and test harnesses emerges quickly.

• •Adoption signals from managed-agent and Agent Platform integrations and any enterprise case studies Google publishes.
• •Independent benchmark replications for UI-control tasks and robustness tests across different OS/browser versions.
• •Third-party tooling for sandboxing, replay-based testing, and human-in-the-loop confirmation workflows that pair with visual actuation.
• •Security analyses showing how effective the adversarial training and the optional safeguards are in real-world prompt-injection scenarios.