OpenAI's Answer to Claude Cowork Is Live. Anthropic's Own Usage Data Says the Real Fight Isn't About Coding.

OpenAI’s ChatGPT Work, launched July 9 alongside GPT-5.6, is built to run inside a company’s existing stack — Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRM platforms, and project trackers — and to stay on a single project for hours, breaking it into steps and completing them with minimal supervision. It ships first to ChatGPT Pro, Enterprise, and Edu users, with Plus and Business following within days, and arrives paired with two related features: Sites, a public beta for turning a ChatGPT session into a shareable web app, and Scheduled Tasks, which let the agent run once, repeat on a schedule, or watch for an event. The framing inside OpenAI’s own announcement and in outside coverage is explicit: this is OpenAI’s answer to Claude Cowork.

That framing raises the question Anthropic already answered with its own telemetry. Anthropic’s usage analysis of 1.2 million Cowork sessions across more than 600,000 organizations found software development is just 8.7% of what people actually do with an agent seat — business process and operations leads at 33.4%, followed by content creation at 16.4%, with sales and revenue operations trailing at 4%. If ChatGPT Work’s CRM and project-tracker integrations pull in a similar mix, the competition between the two products won’t be decided by coding benchmarks; it’ll be decided by which agent handles a reconciled spreadsheet or a drafted proposal with less oversight.

For companies evaluating agent seats rather than picking a single vendor, the practical test isn’t “which agent codes better” — it’s which one integrates cleanly into the tools where the bulk of non-coding work already lives, and whether Anthropic’s own usage data, gathered on a different product, actually predicts how ChatGPT Work gets used once it’s live.