Anthropic Expands Claude Cowork to Mobile and Web, Usage Data Shows Over 90% of Sessions Have Nothing to Do With Coding

What Happened

Anthropic expanded Claude Cowork beyond its original macOS and Windows desktop home, making it available on mobile and web starting Tuesday, according to VentureBeat, The Verge, and ZDNET. The rollout starts with Max plan subscribers, with broader access coming to other Claude tiers in the coming weeks.

Cowork sessions now run in the cloud by default. Previously, the agent ran locally on your machine. Now a task can start on a laptop, continue in the background after the lid closes, and surface notifications on a phone when the AI needs a decision. Scheduled tasks can run entirely in the cloud as long as they only touch connected apps, such as email threads, calendars, and news feeds. Local file access remains desktop-only.

Anthropic is also extending doubled Cowork usage limits through August 5.

The Usage Numbers

Alongside the product launch, Anthropic published data drawn from 1.2 million anonymized Cowork sessions sampled between May 11 and May 31, covering more than 600,000 organizations. The breakdown cuts against the dominant enterprise AI narrative, which has centered almost entirely on coding assistants and developer productivity.

Business process and operations — pulling scattered updates into single reports, building onboarding checklists, reconciling spreadsheets — was the single largest category at 33.4% of all sessions, according to VentureBeat. Content creation and copywriting came second at 16.4%. Together they account for roughly half of all Cowork usage.

Software development accounted for 8.7%. DevOps and infrastructure: 7%. Research and intelligence: 6.4%. Data analysis: 5.8%.

ZDNET's framing put it plainly: more than 90% of Cowork sessions aren't for coding. That's a fundamentally different customer than the one most AI companies have been pitching to investors.

Anthropic's own language for this is "the work around the work" — the administrative connective tissue that spans every role but appears in nobody's job description. Digging through email chains, consolidating status reports, organizing files. Drudgery that consumes hours and produces no core deliverable on its own.

Autonomous AI and Approval Oversight

When an AI agent runs in the cloud while you're in a meeting, it operates without real-time human oversight. Anthropic's safeguard is the notification-and-approval system: Cowork pushes a phone alert when it hits a decision point, and drafted messages aren't sent until the user explicitly approves. No action fires without review.

Skeptics can reasonably ask whether a quick phone tap while distracted in a meeting constitutes meaningful oversight, or whether it's a rubber stamp. ZDNET's reviewer noted discomfort with the idea of granting permissions to an AI from a phone notification. That's a genuine usability and accountability question the industry hasn't resolved. Anthropic has designed a checkpoint system, but whether the checkpoints are substantive in practice depends entirely on user behavior.

Open Source and Frontier AI Models

On Monday, Decagon CEO Jesse Zhang published an argument that open source and frontier AI models aren't really competing — they serve different phases of the same deployment lifecycle. Cheaper open source models absorb mature, well-defined use cases; frontier models keep dominating early-stage discovery and novel deployments.

Data from Vercel's AI gateway dashboard, cited by TechCrunch, supports that frame: DeepSeek now processes just over a third of tokens flowing through Vercel's infrastructure by volume. But Anthropic still accounts for more than half of total AI spend on that same platform. On OpenRouter, DeepSeek V4 Flash leads on raw token volume at 5.3 trillion tokens weekly. The most popular frontier model, Opus 4.8, handles just over 2 trillion — but at roughly 23 times the per-token cost ($1.37 per million versus 6 cents), it likely captures the majority of spending.

The Cowork usage data fits neatly into Zhang's thesis. The "work around the work" category — messy, varied, hard-to-automate administrative tasks — is exactly the kind of early-stage, high-complexity use case where users still reach for the most capable model they can get. Standardized coding tasks can be handed off to lighter models. Drafting a sensitive executive summary from five contradictory status reports probably can't.

What Comes Next

The expansion to additional Claude plan tiers is described as arriving "in the coming weeks," without a specific date. The doubled usage limits run through August 5. The full Cowork feature set, including local file access, remains desktop-only for now.

The unresolved question isn't technical. It's whether Anthropic's notification-based approval model will hold up under real-world use patterns, or whether autonomous cloud sessions will eventually produce an error that reaches a customer before a human sees it. That's the stress test the product hasn't publicly faced yet.