Member-only story
Anthropic just published the most inconvenient dataset in AI. On July 7, the company released usage data from 1.2 million anonymized Claude Cowork sessions across more than 600,000 organizations — and software development, the use case this entire product category was supposedly built for, accounted for just 8.7% of them. Forty-eight hours later, OpenAI shipped ChatGPT Work, a near-exact clone of Cowork, rebuilt its Mac app around it, and buried the chat interface that made ChatGPT famous in a sidebar.
I’ve been running Claude Cowork daily since February, and the 8.7% number matches what I see in my own usage logs almost perfectly. The agent that was marketed as “Claude Code for people who don’t code” turns out to be something else entirely: a machine for what Anthropic now calls “the work around the work.” And the fact that OpenAI reorganized its flagship product around this exact insight, two days after Anthropic published it, tells you where the next twelve months of the AI race are going.
Let me walk through the data, what OpenAI actually shipped, and — because I use one of these tools every day — how to get real value out of the agentic workflows both companies are betting the house on.