Claude Sonnet 4.6 introduces a major leap in AI capabilities with its focus on computer use, marking a shift from traditional API-based tool calling to direct interaction with software through graphical user interfaces. This change represents a fundamental evolution in how AI agents operate. Tool calling, the standard approach in most agentic systems today, relies on predefined APIs. The model generates structured JSON to request a function, which is executed and returns results. While effective, this method is limited to capabilities already exposed as endpoints. If no API exists, the AI cannot act. Computer use, in contrast, allows the AI to interact with software just like a human would—by viewing screenshots, interpreting on-screen elements, clicking, typing, and scrolling. It doesn’t need an API; it works directly with the visual interface. This is the core innovation in Claude Sonnet 4.6. Sonnet 4.6 is an upgrade to Sonnet 4.5, maintaining the same pricing and context window but delivering better performance across the board. In testing, users preferred it over Sonnet 4.5 in about 70% of cases and over Opus 4.5 in 59% of comparisons—remarkable for a model in the Sonnet tier. Key improvements include adaptive thinking, where the model adjusts its reasoning effort based on task complexity, stronger instruction following, and reduced tendency to over-engineer solutions. But the standout feature is computer use. The underlying mechanism is a simple loop. You give the AI a task—like filling out an expense report or searching for flights. It takes a screenshot, analyzes the screen, decides on an action (e.g., moving the mouse to a specific coordinate and clicking), and executes it. The system captures a new screenshot, evaluates the result, and repeats until the task is complete. The AI never sees the app’s internal code or APIs. It only sees pixels, just like a human. It reasons based on visual feedback and adapts dynamically. However, this approach is slow. Each action involves a full round trip: screenshot capture, upload, processing, decision, response, and execution—adding seconds per step. A 20-step task could take 2–3 minutes. Additionally, screenshots are processed as vision tokens, with size limits and added cost. Because of this, Anthropic recommends using computer use for non-time-sensitive tasks like background data gathering, automated testing, or batch processing—not real-time interaction. A smart agent knows when to use the right tool. It uses bash for command-line tasks, a text editor for file reading, and only resorts to screenshots when visual interaction is truly needed. The bottleneck is real, but manageable with thoughtful design. The agent loop is implemented using standard agentic patterns. You define tools—computer, bash, text editor—and send messages. The model responds with tool use requests, which your system executes and returns. The loop continues until completion. Anthropic provides a complete reference implementation via Docker. Running it requires only a few commands. Once launched, you can access the interface at localhost:8080 and watch Claude interact with a real computer—clicking, typing, navigating—without any API wrappers. This demo proves that AI can now use software directly, opening the door to more autonomous, flexible, and human-like automation. While not fast enough for real-time use, it’s a powerful step toward true AI agents that can operate across the full spectrum of digital tasks.