Anthropic Launches New "Compute Usage" Feature for Claude, Sparking Early User Exploration

Recently, the AI company Anthropic introduced a new feature for Claude called "Computer Usage." Although still in the testing phase, this feature has garnered significant attention from early adopters with varying technical expertise, who are actively exploring its capabilities. Users are leveraging Claude for a range of applications, including complex programming tasks, in-depth research, and comprehensive information synthesis.

"Computer Usage" enables Claude to autonomously operate a computer, perform repetitive tasks, and rapidly aggregate data from multiple sources. This groundbreaking functionality has profound implications for future work paradigms.

Claude is equipped with "visual" and autonomous operating capabilities, allowing it to "see" screen content through screenshots, adapt to diverse tasks, and seamlessly switch between different workflows and software applications. It can navigate across multiple screens, applications, and browser tabs, launch programs, move the cursor, click buttons, and input text.

For instance, in a demonstration video, a user instructed Claude to research current AI news stories and provide an overview. Claude proceeded to open a browser, move the cursor to the address bar, type in "Reuters," navigate to the AI section, and repeated the process for The Verge and TechCrunch. Ultimately, the model delivered six trending news stories.

In another example, Anthropic researchers tasked Claude with gathering information about a specific vendor. The model began by taking screenshots, identified missing entries for the vendor, navigated to the Customer Relationship Management (CRM) system, located the company, conducted a search, and successfully matched the information. Subsequently, Claude autonomously transferred the data, filled in required fields, and submitted the vendor form.

Additionally, an Anthropic employee showcased how to use Claude in conjunction with the bash tool (a command language) to download a random dataset, install the open-source machine learning library sklearn, train a classifier on the dataset, and display the results—all within just five minutes.

Notably, the new feature also allows Claude to bypass human verification controls designed to prevent unauthorized access. Some users have reported that their Claude agents are now capable of solving CAPTCHA verifications and successfully logging into ChatGPT.

However, Anthropic researchers have also observed intriguing and anthropomorphic behaviors, such as Claude unexpectedly switching to browsing photos of Yellowstone National Park during a coding demonstration, seemingly mimicking human procrastination.

As the "Computer Usage" feature continues to evolve and improve, Claude's potential applications are expanding, but this advancement may also introduce a series of new challenges and ethical considerations that warrant ongoing attention from the industry and the public.