Claude Cowork Goes Mobile, and Anthropic's Own Data Shows It's Barely Used for Coding

Anthropic launched its Cowork agent on mobile and web this week, letting tasks that once lived entirely on a laptop follow users onto their phones, keep running after they close the app, and wait for approval before anything actually ships. The rollout starts in beta for Max subscribers before reaching other plans. But the more revealing part of Tuesday's announcement wasn't the new surfaces. It was the usage data Anthropic chose to publish alongside them.

The company sampled 1.2 million anonymized Cowork sessions from more than 600,000 organizations, collected between May 11 and May 31, and the breakdown undercuts a lot of what the AI industry has assumed about who actually wants these tools. Software development accounted for just 8.7% of sessions. Business process and operations work, things like pulling scattered updates into a report, building onboarding checklists, reconciling spreadsheets, made up 33.4%, nearly four times as much. Content creation, covering drafts, slide decks, posts, and proposals, came in second at 16.4%. Together those two categories account for roughly half of all Cowork activity. DevOps and infrastructure trailed at 7%, research at 6.4%, data analysis at 5.8%, document processing at 4.1%, and sales operations at 4%. Everything else, including personal assistance, education, and meeting intelligence, stayed under 4% individually.

Anthropic has a name for what dominates: "the work around the work." It's the framing the company uses for tasks that touch almost every job but belong to none of them specifically, the connective labor that keeps projects moving without ever showing up on a resume. It's also a deliberate repositioning. Rather than pitching AI as a replacement for professional expertise, Anthropic is arguing its most valuable current use is clearing away everything that sits between a professional and the actual exercise of that expertise.

What mobile actually changes

The new release adds three capabilities that matter beyond the obvious convenience of a phone app. Sessions now sync across devices, so a task started at a desk can be checked on, and retrieved, from anywhere. More significant is that Cowork can now run entirely in the background with no device online. Users can schedule work in advance, and Claude executes it on its own. Anthropic's example: set up Monday client prep for 6 a.m., and by the time you're at your desk, Claude has worked through email threads, call transcripts, and recent news, built a briefing document, and left a follow-up email drafted but unsent. The third piece is a judgment checkpoint. When Claude hits a decision that needs a human call, it pushes the question to the user's phone. "Nothing ships until you've reviewed and approved it," the company states.

Desktop still has the deepest feature set, including access to local files and the browser. But the web version matters for a different reason: it reaches users in corporate environments where IT departments don't let employees install desktop software, which is a real barrier in a lot of large enterprises. Anthropic also merged its interfaces, so chat and Cowork now share one home screen with projects and outputs persisting across both modes, and it's extending doubled usage limits through August 5 to push adoption during the transition.

The strategy behind chasing non-developers

Anthropic is running two distinct plays at once. Claude Code, its terminal-based agent, is built for and dominated by software engineers doing the actual work of building, debugging, and shipping. Cowork is aimed at the much larger population of people whose jobs don't involve writing code at all. Tellingly, even developers who use Cowork tend not to use it for programming. They use it for the status updates, documentation, and coordination that surround their actual job, the same pattern showing up everywhere else in the data.

Anthropic frames this as people using Cowork to assemble and structure information so they can focus on applying their actual expertise, and offers three examples: a lawyer who lets Cowork handle document formatting and filing while keeping legal judgment for themselves, a hiring manager who has it synthesize interview feedback so they can spend more time actually talking to candidates, and a team lead who has it build the slide deck explaining a decision rather than the decision itself.

The business logic is straightforward. Developer tools are high-profile but serve a narrow slice of the workforce. Cowork targets every knowledge worker with a laptop and a deadline, a market several orders of magnitude larger. There's already evidence the broader enterprise push is working: the Ramp AI Index published in May found 34.4% of firms now pay for Anthropic's services versus 32.3% for OpenAI, the first time Anthropic has pulled ahead, with Claude Code identified as the primary driver.

A busy stretch gets busier

The mobile launch lands in the middle of one of Anthropic's most active periods yet. Last week the company shipped Claude Sonnet 5, which narrows the performance gap with its pricier Opus-class models while undercutting them on cost, introductory pricing of $2 per million input tokens through August 31 before it rises to $3. Sonnet 5 is the model running underneath Cowork, so its gains in reasoning, tool use, and sustained task execution flow directly into Cowork's ability to handle longer, more complicated workflows.

Two weeks before that came Claude Tag, a Slack-native agent built for teams rather than individuals. Where Cowork delegates tasks to one person, Tag operates as a shared Claude identity that an entire channel can interact with over time. Anthropic says 65% of its own product team's code is now generated through its internal version of Tag. Cat Wu, Anthropic's head of product for Claude Code and Cowork, put the distinction to Fortune this way: "Claude Code, Cowork, and chat are very single-player, whereas Claude Tag is built to be interactive and multiplayer." Taken together, Cowork and Tag form a pincer move on enterprise workflows, one capturing individual task delegation across devices, the other embedding Claude directly into team communication.

Security questions haven't gone away

The expansion also follows unresolved security concerns about Cowork's desktop version. On July 1, security firm Armadin, led by Mandiant founder Kevin Mandia, published research describing a full sandbox escape on Windows, using DLL sideloading against the Claude desktop executable to gain access to Cowork's virtual machine service and ultimately bypass network restrictions, as reported by SiliconANGLE. Anthropic responded that the issue didn't qualify as a security vulnerability because exploiting it requires an attacker to already have local code execution on the machine. Armadin's broader point stood regardless: running local virtual machines on nontechnical users' systems creates blind spots that standard endpoint security tools struggle to see into.

Moving to web and mobile changes that calculus somewhat. Those versions run tasks server-side instead of in a local VM, which closes the specific hole Armadin found but raises different questions about how data moves through scheduled background tasks that touch email, calendars, and documents without anyone watching in real time. Anthropic maintains that decisions still come back to the user for approval, but as Cowork takes on increasingly autonomous, multi-step work, the surface area for prompt injection grows with it, a risk Anthropic itself flagged when Cowork originally launched in January.

Geopolitics complicates the enterprise pitch

Anthropic's push into enterprise accounts is colliding with international politics, too. CNBC reported Monday that Alibaba will bar its employees from using Anthropic's tools starting July 10, placing Claude Code on a high-risk software list, a move that followed Anthropic's June letter to the U.S. Senate accusing Alibaba of running what it called the largest known distillation attack against its models. Anthropic is also reportedly closing loopholes that had let Chinese companies reach Claude through third-country entities, part of a broader tightening around U.S. export restrictions.

None of that is slowing Anthropic's infrastructure buildout. Reuters reported Monday that the company signed a 20-year, $19 billion lease with TeraWulf for a Kentucky data center expected to deliver 401 megawatts of computing power once fully operational in 2028. A commitment that size only makes sense if Anthropic expects enterprise demand, not just from engineers but from the much larger population Cowork is chasing, to keep climbing for years.

What the data doesn't show

Anthropic is upfront about the gaps in its own report. The categories describe the type of work being done, not the job title of the person doing it, so there's no standalone bucket for marketing, finance, or HR, those functions likely fold into the business process and operations category, which may inflate its already dominant share. The sample is rate-capped rather than proportional to actual traffic, peak usage hours are somewhat underrepresented, and roughly 5% of sessions were personal rather than work-related. The company also changed its labeling pipeline around May 11, which is why the analysis window starts there instead of covering a longer stretch.

Even with those caveats, the picture is consistent with what Anthropic is building toward. The most valuable application of an AI agent, at least so far, doesn't look like writing code. It looks like turning a mess of emails, transcripts, and spreadsheets into something a person can actually act on, then getting out of the way before anything ships without a human saying yes first.