What Is Claude Cowork? Anthropic’s Agent for Everyday Work, Explained

Claude Cowork is Anthropic’s AI agent for the work that fills most office days: research, reports, spreadsheets, organised files and finished documents rather than lines of code. Give it a goal, grant it access to the folders and tools it needs, and it plans the task, does the work and hands back something ready for review. It launched on desktop in January 2026, and on 7 July it reached web and mobile in beta, which means a task started at a desk now carries on in the cloud while the laptop sits shut in a bag. Two days later, OpenAI answered with ChatGPT Work. The agent race for ordinary office work has properly started, and for once the interesting product is the one that has already been in use for seven months.

Claude Cowork at a glance

What exactly is Claude Cowork?

Anthropic describes Cowork as Claude Code for the rest of your work, and the lineage is the honest way to understand it. Claude Code earned its reputation by letting developers hand over whole programming tasks rather than asking questions one at a time. Cowork applies the same agentic approach to non-technical work. You describe an outcome in plain English, choose which folders and tools Claude can touch, and it gets on with it: reading files, running its own analysis in an isolated environment, working several strands in parallel where that helps, and delivering formatted documents, organised folders or synthesised research at the end.

The control model matters more than the party tricks. Claude shows its plan and each step as it works, waits for approval before anything significant, and cannot delete anything without your sign-off. When a task involves other software, it reaches for direct integrations first, falls back to your browser when it has to, and only takes over your screen as a last resort, because a proper connection is faster and more precise than an AI squinting at pixels.

Chat and Cowork now share one home, so switching between a conversation and a delegated task is a toggle in the message box rather than a separate product to learn.

Which plans include Claude Cowork?

Cowork is included with every paid Claude plan, from Pro upwards, through the desktop app on macOS and Windows. It started in January as a Max-only preview, demand pushed Anthropic to open it to Pro subscribers within days, and it left preview to become generally available on 9 April. There is no free-tier version, which remains one of the clearest dividing lines between Anthropic’s launch and OpenAI’s response. The web and mobile beta is narrower for now: it opened with the Max plan on 7 July and is still working through the other paid tiers.

There is no separate price either. Usage draws on your plan’s allowance and scales with the size and complexity of each task, so a quick file tidy costs a fraction of a research synthesis that runs for an hour. The promotional doubling of limits that marked the July launch ended on 5 August, so the metering you see now is the normal arrangement. The models doing the work have quietly improved as well: since 24 July, Max accounts default to Claude Opus 5, which our review covers along with one caveat on factual reliability, while Sonnet 5 carries the other plans with near-flagship capability at everyday prices.

What changed on 7 July

The web and mobile expansion sounds like a small platform announcement and is actually a change in kind. Cowork sessions now run remotely on Anthropic’s servers, in an isolated temporary environment created for each session, rather than depending on your machine staying awake. Start a task from claude.ai or the mobile app, close the laptop, and the work carries on. Scheduled tasks run with no device online at all.

When Claude finishes or needs a decision, your phone gets a notification, and because sessions follow your account rather than your hardware, you can open the same task from any surface to check progress, answer a question or redirect the work. The desktop app remains the full experience for now, since local file access, browser control, computer use, live artifacts and plugins that include local MCP servers all still need it.

What Anthropic’s usage data shows

The most useful thing Anthropic published alongside the expansion was not a feature list but a number. From an analysis of 1.2 million anonymised Cowork sessions across more than 600,000 organisations in May 2026, over ninety per cent of the work people delegate to the agent has nothing to do with software development. Business operations tasks lead at 33.4 per cent, content creation and copywriting come second at 16.4 per cent, and actual coding trails at 8.7 per cent.

Read that again with a marketer’s eyes. The technology was built for developers, and the moment it was packaged for everyone else, everyone else became the market. It also explains the calendar: OpenAI shipped ChatGPT Work two days after Cowork went multi-platform, built on its own coding agent in exactly the same way. The tools that spent 2025 fighting over programmers are now fighting over the rest of the office.

Claude Cowork vs ChatGPT Work

We have covered OpenAI’s launch in full, so the short version here. Both products take a brief and return finished work, both grew out of coding agents, and both meter usage by the size and complexity of what you ask. The practical differences today come down to access and maturity. ChatGPT Work arrived across web, mobile and desktop in one go, and its desktop app covers every plan including Free, which makes it the cheaper door to walk through. Cowork requires a paid plan and its web and mobile access is still a staged beta, but the desktop product has seven months of production use behind it, along with the admin controls, usage analytics and OpenTelemetry monitoring that Team and Enterprise deployments actually need. Neither has a meaningful capability moat over the other yet. If your organisation already runs on Claude or on ChatGPT, the sensible pilot is the agent attached to the subscription you have.

Why this matters for marketing and business teams

Content creation being the second-largest use of Cowork tells you where the immediate value sits for agencies and marketing teams. The good early workloads share a shape: recurring reports assembled from the same sources every week, research synthesised from a pile of documents, files reorganised and renamed to a convention, first drafts built from briefs and source material. Work with a defined outcome and existing raw material, in other words, where the agent’s job is assembly and judgement rather than invention.

The gap between a demo and a dependable result is almost always on the customer’s side of the fence, though. An agent is only as useful as the context you can hand it, and that means documented processes, structured content, clean data and briefs that would survive contact with a new employee. We made the same argument when Fable 5 returned, and the advice has not changed with the interface: the businesses getting value from agents are the ones that prepared their information before the tools arrived.

Two cautions belong in any pilot plan. First, agents that read the open web can be fed malicious instructions hidden in pages and documents, so Anthropic’s own guidance is to scope folder access tightly, keep an eye on write actions and extend internet access only to sites you trust. Second, some enterprise plumbing is still catching up: Cowork activity is not yet captured in compliance audit logs, conversation history is stored locally on each user’s machine, and memory works only inside projects. None of that blocks a pilot. All of it belongs in the conversation with whoever signs off your data policies.