Google has rolled out Gemini 2.5 (Computer Use), a new AI model that can navigate the web in ways similar to a human user—clicking, scrolling, filling forms, hovering, typing, and interacting with user interfaces.
Key Features & Capabilities
- Gemini 2.5 is built on Google’s “Gemini 2.5 Pro” architecture and is designed to perform web tasks that require complex navigation.
- The model supports 13 discrete actions (such as clicking, dragging, opening menus, inputting text) and controls a virtual browser to interact with web pages.
- While it doesn’t yet manage desktop-level operations, it already enables web and mobile-based workflows.
- Google has demonstrated use cases in which Gemini reorganizes sticky-note boards, fills in forms, and navigates dropdowns, all driven by natural language commands.
- According to Google, Gemini 2.5 outperforms other comparable systems on multiple benchmarking tests for web and mobile navigation.
Internal & Developer Use Cases
- Within Google, teams are applying Gemini 2.5 for tasks like UI testing, where automating interactions can speed up software verification.
- Variants of this model also underpin features in Gemini’s agentic mode in Google Search, as well as in tools like Firebase Testing Agent and an internal initiative known as Project Mariner.
- Developers can access the model via Google AI Studio and Vertex AI, enabling integration into their own applications or workflows.
Some Limitations & Caution
- The model is currently limited to web browsing and cannot yet operate on the desktop or perform tasks outside the browser environment.
- Because demonstrations are often sped up, real-world latency and responsiveness may vary.
- The set of supported actions, though broad, is still fixed—there are limits on what it can do in novel or highly custom web interfaces.
Outlook & Implications
Gemini 2.5 marks a step forward toward more autonomous, interactive AI agents that don’t just respond to queries but perform tasks online. As it develops, this kind of capability could transform how AI handles web automation, virtual assistance, data entry, research, and more.
