Mountain View, California · 7 October 2025Google DeepMind released the Gemini 2.5 Computer Use model in preview on October 7, 2025, making it available to developers via the Gemini API in Google AI Studio and Vertex AI. The model — a specialized variant of Gemini 2.5 Pro — is designed to enable AI agents to interact directly with graphical user interfaces (GUIs) such as web browsers and mobile applications by perceiving on-screen content and performing actions including clicking, typing, and scrolling.The release marked a significant step in the evolution of "agentic AI" — systems capable of performing multi-step actions autonomously rather than merely generating text in response to prompts.## How Computer Use WorksGemini 2.5 Computer Use operates through what Google describes as a "computer use" loop. The mechanism is conceptually straightforward but technically demanding:1. Perception: The model receives a task description and a screenshot of the current user interface.2. Action generation: The model analyses the screenshot and generates structured function calls — such as click_at(x, y), type(text), or scroll(direction) — that correspond to specific actions on the screen.3. Execution: A client application executes these actions on the actual computer or virtual machine.4. Feedback: A new screenshot is captured and fed back to the model, which assesses the result and determines the next action.5. Iteration: This loop continues until the task is complete.> [!STAT] The model processes screenshots of graphical interfaces and generates structured function calls (click_at, type, scroll) to perform tasks — enabling AI agents to operate any software with a visual interface, without requiring dedicated APIs or integrations.## Why This Matters: The Universal Interface ProblemUntil now, AI agents that needed to interact with software typically required dedicated APIs — programmatic interfaces that expose specific functions in a structured way. This works well for major platforms that offer APIs (Google Workspace, Slack, Salesforce), but it leaves a vast landscape of software inaccessible to automated AI agents.Many enterprise applications, legacy systems, and specialized tools have no API at all. They are designed to be operated by humans clicking buttons, filling forms, and navigating menus. The Computer Use model addresses this gap by treating the graphical interface itself as the API — if a human can see it and click it, the AI agent can too.> [!INSIGHT] The significance of computer use models extends beyond convenience. They represent a shift from AI that assists humans (generating text, answering questions) to AI that acts on behalf of humans (completing tasks across multiple applications). This is the foundational capability for the broader vision of agentic AI that dominated the industry's roadmap throughout 2025.## The Competitive LandscapeGoogle's release came amid intensifying competition in the agentic AI space. By late 2025, multiple major AI labs were pursuing computer use capabilities:| Provider | Approach | Status (Oct 2025) ||---|---|---|| Google DeepMind | Gemini 2.5 Computer Use (specialized model) | Released in preview || Anthropic | Claude Computer Use | Previously released || OpenAI | GPT-5.1 with adaptive reasoning | Released November 2025 || Chinese labs | DeepSeek, Qwen, Kimi | Closing capability gap |The State of AI Report 2025, published in October, noted that while OpenAI maintained a narrow lead in overall model capability, international competition had intensified significantly, with models from Chinese labs closing the gap in reasoning and coding capabilities.## From Chat to Action: The Agentic ShiftThe release of Gemini 2.5 Computer Use was emblematic of a broader industry transition that defined late 2025. According to surveys from this period — including the State of AI Report 2025 and McKinsey's State of AI findings — approximately 80% of AI-native companies were investing in autonomous systems capable of executing multi-step actions on behalf of users.The shift from conversational chat to autonomous action had several drivers:- Maturation of multimodal models: AI systems that can process images (screenshots) alongside text are a prerequisite for computer use capabilities.- Improved reasoning: Models capable of multi-step reasoning can plan sequences of actions to achieve complex goals.- Enterprise demand: Businesses sought AI that could automate workflows across multiple applications, not just answer questions.## Technical Considerations and LimitationsThe computer use approach, while powerful, comes with inherent challenges:- Speed: The perception-action loop requires capturing screenshots, sending them to the model, and receiving actions back — a process that is inherently slower than direct API calls.- Reliability: GUIs change frequently (new layouts, pop-ups, loading states), and models must handle these variations robustly.- Security: Giving AI agents the ability to click and type on real computers raises significant security considerations, requiring sandboxed environments and careful access controls.- Cost: Each loop iteration involves processing an image through a large model, which is more computationally expensive than text-only interactions.> [!NOTICE] As of August 2026, Google has integrated computer use capabilities directly into newer model families (Gemini 3 Pro and Flash series), making the standalone Gemini 2.5 Computer Use preview model a legacy component. The underlying capability — AI agents operating graphical interfaces — has become a standard feature rather than a specialized offering.## The Broader TrajectoryThe release of Gemini 2.5 Computer Use in October 2025 represented a specific moment in a larger arc: the transformation of AI from a tool that generates content to a tool that takes action. By enabling AI to operate any software through its visual interface, Google and its competitors were building toward a future where AI agents could handle complex, multi-application workflows — booking flights, filling out expense reports, navigating enterprise software — without human intervention at each step.Whether this future arrives quickly or gradually depends on solving the reliability, security, and cost challenges inherent in the computer use paradigm. But the direction is clear: the interface between AI and the digital world is expanding from text to the full visual landscape of human-computer interaction.## Sources- Google DeepMind, "Introducing Gemini Computer Use," blog.google, October 7, 2025- InfoQ, "Gemini Computer Use: Google Brings Agentic UI Control to Developers," October 2025- DataCamp, "Gemini 2.5 Computer Use Guide," datacamp.com- State of AI Report 2025, stateof.ai, October 2025- McKinsey, "The State of AI," mckinsey.com, November 2025- Google AI documentation, ai.google.dev/gemini-api/docs/computer-use
FIRAT Editorial Board
Institutional Research Desk · Foresight Institute of Research and Translation
The collective editorial and research translation board of FIRAT, synthesising peer-reviewed evidence, policy briefs, and division milestones across our seven foundational research pillars.
