OpenAI hits back at Anthropic with ChatGPT Work
A day after GPT-5.6, OpenAI unveils ChatGPT Work, an agent that builds sites, reports and slides on its own and works for hours on one task.
- On July 9, OpenAI unveiled ChatGPT Work, an agent that builds websites, reports and presentations on its own.
- It runs on the new GPT-5.6 model, which ships in three flavours: Sol, Luna and Terra.
- The tool is a direct answer to Anthropic's Claude Cowork, released in January 2026.
- The rollout starts with Pro, Enterprise and Edu accounts, then reaches Plus and Business.
Do you want an assistant that does not just write text, but actually finishes the whole task? OpenAI thinks the answer is yes, so on July 9 it unveiled ChatGPT Work, a workplace agent that assembles documents and sites for you.
The timing is no accident. A day earlier came the new GPT-5.6 model, and ChatGPT Work is the first big thing to use it. The message is clear. OpenAI does not want to stay a chat tool, it wants to be a colleague that takes work off your plate.
What exactly does ChatGPT Work do?
Unlike an ordinary chat, this agent does not stop at a single reply. It connects to your apps and files, then works for hours on one complex task. Ask for a report and it writes it. Need a spreadsheet pulling data from several sources and it assembles it. A deck for a meeting, a site for a small business, all of that is in the job description.
The key word is hours. Today’s AI tools work in seconds and hand you something you then polish by hand. OpenAI is now pushing the idea that an agent can work for a long time, on its own, on a project with many steps. That is a different way of working from anything we have tried in daily use so far.
Under it all sits GPT-5.6, and it does not come in one version but in three. Sol is the most powerful, for the hardest jobs. Luna is optimised for speed, when you need an answer right away. Terra is the middle ground, an everyday balance of power and speed. Sam Altman claims the model is 54 percent more token efficient at agentic coding, which in plain terms means cheaper work for companies watching every dollar they spend on AI.
On top of that, OpenAI is cleaning house. The Codex coding tool is folded into a single desktop app, and the Atlas browser is being wound down. Fewer scattered products, one agent at the centre.
Why now, and what does it mean for work?
Because Anthropic has not been asleep. Its Claude Cowork shipped in January 2026 and was the first to seriously push the idea of an AI colleague that takes on tasks. ChatGPT Work is a direct answer to it. The two biggest AI companies are now openly fighting over the same ground: the workplace of the future.
The rollout does not start for everyone at once. Pro, Enterprise and Edu users get it first, then it widens to Plus and Business accounts. It works on web, mobile and desktop. The logic is plain. First the companies and schools that pay more, then the rest.
One detail from the background is worth remembering. The GPT-5.6 launch was delayed earlier, at the request of the US government, over safety fears. A model that now works for hours on its own clearly raised enough caution to be discussed at the state level. That is no small thing.
Our advice is not to rush the whole job onto the agent on day one. Tools like this shine on repetitive tasks with clear rules, and stumble where judgment is needed. Let it draft the first version of a report or a spreadsheet, then check it. The time saved is real, but the checking stays with you.
Would you let an AI agent work on your task for hours on its own, or would you still want to see every step?
