Anthropic unifies Claude chat and Cowork, adding Docs and Slides to one work surface

What the unified Claude can do, according to Anthropic. The new experience supports quick answers, multi-step tasks, web search, reading user files, code execution, creation of documents/spreadsheets/presentations, connected-app access, scheduled work, mobile check-ins, skills, plugins, projects, memory, and artifacts. Anthropic says more involved tasks can keep running in the cloud after the user closes the page or laptop, but tasks that need files or apps on the user’s computer require Claude Desktop to be open. Claude Cowork and chat are one Claude | Claude Help Center

Access. The unified experience is rolling out gradually first to Pro and Max users on web, desktop, and mobile; more plans follow later, and Enterprise admins get at least 30 days’ notice before changes. Anthropic’s Help Center says that once a Pro/Max account receives the new experience, it cannot switch back to separate Chat and Cowork options. Claude Cowork and chat are one Claude | Claude Help Center

Docs access has a nuance. The TechCrunch excerpt says the new features roll out to Pro and Max first and later to Free and Team. Reuters similarly reports Pro and Max first, then Team and Free. Anthropic’s own Docs Help Center, however, says Claude Docs is available in beta on Pro, Max, Team, and Enterprise, not Free, on by default for Pro/Max/Team, off by default for Enterprise, and unavailable for organizations using CMEK, ZDR, or HIPAA-ready configurations. The safest reading is that “one Claude” may eventually reach Free, while Docs/Slides availability on Free is not clearly supported by Anthropic’s current Docs documentation. Anthropic to fold Claude AI features into one interface, launches document tools By Reuters

Exact model/version. The merger announcement does not specify a required Claude model or a new model version. Anthropic’s Help Center says the model picker remains where it was, and separate release notes show Anthropic had launched Claude Fable 5.1 and Mythos 5.1 earlier in September, but the “one Claude” change should be treated as a product-surface change unless Anthropic states otherwise. Claude Cowork and chat are one Claude | Claude Help Center

The documented architecture is best understood as a single conversational front end over multiple execution modes. Anthropic does not publish the routing policy, classifier, or planner that decides whether a prompt remains a quick chat or becomes an agentic task. It does document that cloud Cowork sessions run the agent loop and code execution on Anthropic servers, with session files saved to the member’s Claude account; existing desktop deployments can still use local execution. Claude Cowork architecture overview | Claude Help Center

For cloud sessions, Anthropic says each session runs in an isolated, temporary sandbox, with no default access to private/internal/link-local/cloud-metadata addresses; egress passes through a mandatory proxy; credentials are short-lived and session scoped; connector authorization tokens do not enter the sandbox; and stored records are tenant scoped. When a cloud session needs local files or browser access, it reaches the device through Claude Desktop over an Anthropic-brokered connection, limited to connected folders and only while the desktop app is online. Claude Cowork architecture overview | Claude Help Center

For local sessions, Anthropic says the agent loop runs natively on the user device, while shell commands and code execute in a dedicated Linux VM isolated through Apple Virtualization.framework on macOS or Hyper-V on Windows. That architecture matters for enterprise security teams because the merged UI hides product-mode selection from the user but does not eliminate execution-mode differences, data-location differences, or permission differences. Claude Cowork architecture overview | Claude Help Center

For developers building internal analogues, the implementation implications are clear even though Anthropic’s router is opaque:

1. Expose execution state. Users need to know when a request has moved from chat to a long-running tool-using agent.
2. Separate read and write tools. Anthropic’s safety guide explicitly distinguishes read tools from write tools and says write tools carry more risk. Use Claude Cowork safely | Claude Help Center
3. Instrument agent traces. Anthropic says Cowork via Claude, desktop, and mobile is captured in Compliance API, and Team/Enterprise admins can stream Cowork events via OpenTelemetry. Claude Cowork architecture overview | Claude Help Center
4. Treat untrusted content as adversarial. Academic benchmarks such as AgentDojo and WASP show that tool-using and web agents remain vulnerable to indirect prompt injection even in realistic task settings. AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents | alphaXiv

There is no public reproducibility package for evaluating Anthropic’s new routing behavior. A credible evaluation would need a task suite spanning quick Q&A, long research, doc generation, spreadsheet/file edits, web-browsing tasks, connected-app tasks, and adversarial external content; metrics should include route appropriateness, task completion, token/usage cost, permission prompts, hidden harmful actions, recoverability, and user comprehension.

The move fits a broader industry shift from discrete chatbots toward general AI work surfaces: one interface that can answer, browse, manipulate files, call connectors, create artifacts, and continue work asynchronously. Reuters frames Anthropic’s move as part of a race with OpenAI for enterprise and non-coder AI work tools; The Verge frames Docs and Slides as narrowing the gap with Google’s productivity suite; Axios emphasizes the competitive pressure on Microsoft Office-style workflows. Anthropic to fold Claude AI features into one interface, launches document tools By Reuters

Simon Willison’s practitioner read is more cautious: he interprets the change as Claude becoming “a general agent,” but also notes that understanding the practical boundaries among Claude, Cowork, and Claude Code will still take work. That is a useful warning for buyers: product simplification at the UI layer can make conceptual boundaries less visible, not necessarily simpler in operational terms. Claude Cowork and chat are now one Claude

Anthropic’s own safety guidance is unusually explicit that agentic Claude has “real capabilities”: reading files, browsing, running code, and using apps. It warns that computer use has no sandbox between Claude and what is on the screen, that web content is a primary prompt-injection vector, and that users remain responsible for actions Claude takes on their behalf, including messages, purchases, data changes, and scheduled tasks. Use Claude Cowork safely | Claude Help Center

The Docs feature also has beta limitations: no version history, no comment-only access level, charts/diagrams do not auto-refresh, Team/Enterprise docs cannot yet be shared outside the organization, some regulated configurations are unsupported, and activity inside a doc—edits and comments—is not yet recorded in the Compliance API. Get started with Claude Docs | Claude Help Center

The main contested or unresolved point is availability, especially around Free-tier access to Docs/Slides versus access to the unified interface. Independent reports say Free comes later, but Anthropic’s current Docs page says Docs is not available on Free. Treat plan availability as still moving during rollout. Anthropic to fold Claude AI features into one interface, launches document tools By Reuters

For business leaders, the upside is reduced workflow fragmentation: employees can start from one Claude prompt and end with a doc, deck, spreadsheet, design, or recurring task. That may increase adoption and reduce “tool-choice overhead.” For IT, however, the governance question shifts from “Who has Cowork?” to “Which conversations can become agentic, access which connectors, and take which actions?” Anthropic’s Enterprise notice period and admin controls are therefore more important than the UI change itself. Claude Cowork and chat are one Claude | Claude Help Center

For developers and researchers, the launch underscores that routing, observability, and permissioning are now first-class AI product problems. The model’s raw capability is only one layer; the agent runtime, sandbox, connector model, logging coverage, memory behavior, and human-approval defaults determine whether the system is usable and governable.

For procurement teams, require pilots that measure: task success, rework rate, usage consumption, approval fatigue, data exposure through connectors, audit-log completeness, incident response paths, and employee understanding of when Claude is acting versus answering.