The spreadsheet moment for AI

ON JULY 9TH, alongside the launch of ChatGPT Work, OpenAI merged its standalone Codex application into a rebuilt ChatGPT desktop app. Chat, Work and Codex now sit on every plan, including the free one. Codex was a developer's tool, gated behind a subscription and a tolerance for the terminal; ChatGPT has something near a billion weekly users, most of whom have never opened one. The capability did not change last week; the doorway did.

ChatGPT Work, running on the GPT-5.6 family released the same morning, takes an outcome rather than a question: it gathers context across a user's connected apps, breaks a goal into steps, works for hours while the user is elsewhere, and returns finished spreadsheets, slides, documents and shareable web apps. The coverage has scored it against Anthropic's Claude Cowork, which reached general users first. Cowork already ran the experiment, and published the result.

Terminal velocity

Two days before OpenAI's launch, Anthropic released a study of 1.2 million anonymized Cowork sessions from more than 600,000 organizations, sampled across the last three weeks of May. Cowork descends from Claude Code, an agent built for the terminal and sold to engineers; Anthropic built it, the company says, to extend those capabilities to the chat interface people were already using. Software development accounted for 8.7% of sessions. Business process and operations led at 33.4% — reconciling spreadsheets, consolidating scattered updates into a single report — with content creation at 16.4%, research at 6.4% and sales operations at 4%. Anthropic calls the category "the work around the work." Same machinery as Claude Code, minus the shell, and nine sessions in ten had nothing to do with code.