On Monday, March 24, 2026, Anthropic announced major upgrades to Claude Code and Claude Cowork. The core breakthrough is the integration of a "computer use" capability: the AI is no longer limited to tool calls or API interfaces but can directly view the screen, move the mouse, click buttons and type on the keyboard, enabling autonomous control of the entire computer system.
This update is seen within the industry as the most aggressive capability leap to date for the Claude family, marking a shift from "conversational advice" to "real execution." The upgrade comes amid the viral rise of the open‑source AI agent OpenClaw (launched by developer Peter Steinberger, formerly named Clawdbot before a trademark‑driven rename). OpenClaw, which can remotely control a desktop via mobile messages, has quickly become popular worldwide (notably in China) as a tool for automating everyday tasks. Anthropic's move is widely viewed as a strategic response directly targeting OpenClaw's momentum.
Claude Code was originally positioned as a developer‑focused agent capable of handling complex coding, debugging and multi‑step workflows; Claude Cowork, released in January 2026 as a "consumer" version, is designed for non‑technical users to manage file organization, report generation and data extraction. After this upgrade, both can—subject to user authorization—open files, browse web pages, run development tools and even continue tasks when the user is away from the computer (via a new Dispatch mobile feature that lets users issue commands from their phones).
How it works: moving beyond tool calls to simulate human GUI actions
Unlike traditional AI that relies on predefined APIs or plugins, the new "computer use" employs a vision‑plus‑action loop. Claude first captures screenshots of the current desktop (using its visual model), then reasons about the next action: planning the task, locating buttons, moving the mouse, clicking or typing. The process forms a closed loop—the model re‑screens to verify results after actions until the task is complete.
This is essentially the operationalization of agentic AI: the model no longer passively responds to prompts but actively formulates multi‑step plans, handles exceptions and switches seamlessly across the local file system, browser and development environment. Anthropic says this capability builds on the early computer use beta of Claude 3.5 Sonnet from October 2024, but is now embedded directly in Claude Code/Cowork with no extra setup—plug‑and‑play (currently macOS only, with potential Windows support later).
Technically, it combines very large context windows (million‑token scale), planning tuned by reinforcement learning and real‑time visual understanding to achieve a qualitative change from "thinking" to "doing." Developer tests show Claude can independently complete complex workflows such as organizing folders, extracting screenshot data into spreadsheets, and searching and summarizing browser results, with users only needing to grant initial authorization.
Cost: subscription‑driven, advanced use can be expensive
The upgrade is released as a "research preview" and is available for free trial to Claude Pro and Max subscribers. Pro plans start at roughly $20 per month for standard usage; Max plans are tiered—$100/month (5x usage) or $200/month (20x usage) for heavy users. When billed separately, Claude Code can cost some advanced developers over $200 per month depending on task complexity and token consumption.
API developers who integrate computer use themselves are still charged by token (Sonnet‑series input about $3 per million tokens, output $15 per million tokens; Opus is higher), but long‑running tasks that require repeated screenshots and actions can cause costs to accumulate quickly. Anthropic says it plans to optimize pricing to balance accessibility and compute costs. For now, free users cannot access full computer‑control features, reflecting Anthropic’s commercial stance that advanced capabilities are gated by paid tiers.
Security: permission‑first design, but risks remain
Anthropic adopts a "permissions‑first" approach: Claude must request application access permissions before executing actions, and users can interrupt operations at any time; the system includes prompt‑injection detection, vulnerability scanning and recommends running in sandboxed environments. The company states that 0.8% of operations may be "irreversible" (for example, sending an email), while 73% of scenarios still retain a human‑in‑the‑loop for supervision.
Nevertheless, risks persist: malicious prompt injections could induce the AI to perform harmful actions; deceptive on‑screen content may mislead the model; full access to the file system could lead to data exposure or accidental deletion. Anthropic recommends granting only necessary permissions and auditing logs regularly. Compared with OpenClaw's open‑source flexibility, Claude's implementation emphasizes enterprise‑grade compliance but sacrifices some of the plug‑and‑play spontaneity. Early tests have already prompted security experts to warn that a hijacked AI agent could become an "insider threat."
Impact: a dawning era of autonomous work and accelerated industry reshaping
For ordinary users, this is arguably a productivity revolution. Claude Cowork can automatically manage email triage, draft reports and run analyses while the user is away; developers can use Claude Code for unattended coding and deployments. Firms such as Gartner predict agentic AI could boost knowledge‑worker productivity by 30%–50%, propelling an era of "autonomous work" where AI is not just an assistant but a digital colleague.
At the enterprise level, the implications are even deeper. Sectors like finance, law and marketing have begun pilots: firms including Goldman Sachs are reportedly using Claude agents for accounting and compliance workflows to reduce human labor. SaaS incumbents (e.g., Salesforce, Workday) face pressure as traditional software may be bypassed by AI‑native workflows—but new opportunities also arise, such as demand for security monitoring and agent orchestration platforms.
Competition in the agent space is heating up. OpenClaw's open model put Anthropic on the defensive; this upgrade is seen as a counterpunch, but it also highlights common challenges across the agent sector: cost, reliability and safety. In the long run, this accelerates the shift from "chat AI" to "execution AI" and could reshape employment—routine office tasks may be automated while creative and supervisory roles become more valuable.
Conclusion: the autonomous‑work era is at the doorstep, but caution is needed
Anthropic's upgrade is not only a timely response to OpenClaw's popularity but a key step toward embodied intelligence and real‑world interaction. It shows the technology can "take action," but commercialization, security and ethical issues still require iteration. The autonomous‑work era will not arrive overnight but will emerge through gradual authorization and continuous refinement. For users, now is a prime time to experiment—start with simple tasks and progressively entrust AI with fuller agency.
Anthropic pledges rapid iteration (expanding platform support, reducing latency) and calls for industry‑wide standards on agent safety. In any case, Claude Code and Cowork's "computer takeover" has brought science fiction closer to reality: AI is no longer just a chat tool but can truly act as a working "colleague."
▶️ Enjoyed this piece? Creating it takes real effort!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com