Anthropic puts Claude Cowork on your phone, and the agent race follows
Anthropic's Claude Cowork AI agent now runs on phone and web for Max subscribers. OpenAI's Codex heads the same way, and the contest is for your lock screen.
Anthropic's Claude Cowork AI agent now runs on phone and web for Max subscribers. OpenAI's Codex heads the same way, and the contest is for your lock screen.
Anthropic's Claude Cowork, an AI agent that started as a desktop helper for general knowledge work, now runs on phones and web browsers for the company's Max subscribers, beginning Tuesday. The expansion sharpens the contest over who owns the phone screen where the small chores of office life actually happen.
Cowork's own usage does not match its original pitch. Anthropic's study, as summarized by TechCrunch's Rebecca Bellan, covered roughly 1.2 million anonymized Cowork sessions from the last two weeks of May, drawn from more than 600,000 organizations. The dominant use case was not software development (8.7 percent). It was business process at 33.4 percent, the catchall for briefs, follow-ups, prep notes, and inbox triage. Content creation ran a distant second at 16.4 percent. The typical Cowork session looks more like a junior analyst's task list than a software engineer's pull request.
The product's framing says one thing. The usage data says another. Anthropic has marketed Cowork as "working around the work," language that positions the agent as a coordinator rather than a doer. The published numbers suggest the opposite: users are handing Cowork the actual work, not just the meta-work around it. Both readings can be true at once, but enterprise buyers should read the data as a reason to ask a sharper question: is this agent a sidekick, or is it being asked to do the job?
The mobile and web versions add an always-on layer the desktop app could not. A task can start on a laptop, push status updates to a phone, and finish elsewhere, even with the laptop closed. Anthropic is selling that continuity as the core upgrade, and it is genuinely new for a general-knowledge-work agent: most office-worker AI tools still assume one screen and an active session. The Max-only gating is the obvious commercial hedge. This is a paid-tier feature, and Anthropic has not disclosed any plan to bring it down to free or lower plans.
Codex began as a developer tool but has been pulled toward reports, spreadsheets, presentations, research, and data analysis by enterprise users who wanted the same agent for everything else on their plate. Anthropic and OpenAI are converging on the same target from opposite starting points: Anthropic down from general work, OpenAI up from software. Both labs have already started laying down the always-on surface. Anthropic recently launched Claude Tag, a persistent Claude that lives in Slack. OpenAI has been pushing Codex into ChatGPT, the standalone app, and the desktop. The phone is where the user already is, and that is where the contest will be settled.
The early Cowork data does not settle whether either agent does the work well; it only confirms the volume is there. The Anthropic study is company-issued and aggregated, and the sample stops in late May, before the mobile rollout, so it tells a customer-interest story rather than a quality story. What the data does show is which doors the agent is being asked to walk through, and those doors line up with the screen a knowledge worker checks most often. For most office workers, that screen is the lock screen, and the practical question is which AI agent they will reach for first when the next follow-up pings. Two things will sharpen the picture next: a fresh Anthropic usage study that stretches past the Max tier, or a pricing shift that brings mobile and web down to a broader plan.