Introducing computer use in Gemini 3.5 Flash

Google has integrated computer-use capabilities directly into its Gemini 3.5 Flash model, a move that lets developers build custom agents that can see, reason, and act across browser, mobile, and desktop environments [1]. The feature was previously available only as a standalone Gemini 2.5 computer-use model but is now a native tool within the main Flash variant [1]. Developers can access the capability through the Gemini API and the Gemini Enterprise Agent Platform [1]. Google says the integration improves performance on long-horizon enterprise automation tasks such as continuous software testing and knowledge work across professional applications [1]. Gemini is a family of multimodal large language models developed by Google DeepMind, succeeding LaMDA and PaLM 2, and was first announced in December 2023 [2][3]. The Flash tier is positioned as a cost-effective, high-throughput variant within the broader Gemini lineup, which also includes Nano, Pro, and Ultra versions [2]. The 1.5 and 3 model generations introduced extended context windows capable of analyzing entire codebases or long-form videos in a single prompt [2]. In a demonstration, 3.5 Flash used its computer-use tool to analyze the Gemini application and return a categorized list of features [1]. It also audited its own documentation for accessibility issues [1]. The release comes as agentic browsers—web browsers that can navigate pages or fill out forms on a user’s behalf—gain traction. Several such browsers, including ChatGPT Atlas, Comet, and Dia, emerged in 2025 [5]. Established browsers like Chrome and Edge have also added AI features, with Chrome integrating the Gemini chatbot for U.S. desktop users [5]. To address prompt-injection risks for agents operating in live environments, Google applied targeted adversarial training to the computer-use feature in 3.5 Flash [1]. The company is also releasing two optional enterprise safeguards: one that requires explicit user confirmation for sensitive or irreversible actions, and another that automatically stops tasks if an indirect prompt injection is detected [1]. Google recommends that developers combine these features with secure sandboxing, human-in-the-loop verification, and strict access controls [1]. A demo environment hosted by Browserbase is available for testing, and reference implementations are provided through the Gemini API and Enterprise Agent Platform [1].