On 9 July 2026, OpenAI launched ChatGPT Work, an autonomous agent that lives inside ChatGPT and is pitched, quite openly, as software that can do parts of your job. According to reporting from Bloomberg and Forbes, it takes an outcome you describe, gathers context across your connected apps and files, breaks the goal into smaller steps, and completes them on its own, staying with a complex project for hours before handing back finished work.
What ChatGPT Work actually does
The pitch is a shift from answering questions to delivering results. Rather than returning text you then act on, ChatGPT Work is described as producing finished artefacts: spreadsheets, slide decks, reports, documents, and even working web apps. It pulls the context it needs from the apps and files you connect, plans the task, and executes the steps itself. OpenAI says it will be available across ChatGPT on web, mobile and desktop, rolling out first to Pro, Enterprise and Edu users and expanding to Plus and Business users over the following days.
The engine: GPT-5.6, in three flavours
ChatGPT Work is powered by GPT-5.6, which OpenAI released in three variants according to the reporting: Sol, the most powerful; Luna, optimised for speed; and Terra, tuned to balance performance and efficiency for everyday use. The agent layer also folds in OpenAI's Codex coding technology, which is what lets it generate and run code as part of a larger task rather than just suggesting snippets.
Part of a bigger product reshuffle
The launch does not stand alone. It ships as OpenAI merges its separate Codex app into a single ChatGPT desktop app, and as the company begins sunsetting its Atlas browser. In other words, OpenAI is consolidating its tools around one agentic ChatGPT rather than a spread of separate products.
The workplace-agent race
ChatGPT Work lands months after Anthropic shipped Claude Cowork, and the framing from all sides is the same: the competition has moved from chatbots to autonomous agents that execute multi-step work with minimal human input. Google, Anthropic and OpenAI are now openly racing to own the "agent that does the task" layer, not just the "model that answers" one. For anyone choosing AI tools, that means the interesting question is shifting from which model is smartest to which agent you trust to act on your behalf.
What it means, honestly
An agent that works "for hours" and ships finished output is genuinely useful for well-scoped, repetitive multi-step jobs: assembling a report from known sources, building a first-draft spreadsheet, or scaffolding a web app. It is worth being clear-eyed about the limits, though. Autonomous agents can be confidently wrong, and the longer they run unattended, the more a small early mistake can compound through later steps. Anything an agent ships still needs a human review before it is trusted, especially for anything customer-facing or financial.
There is also a privacy and security dimension that is easy to overlook in the excitement. An agent that "gathers context across your connected apps" is, by definition, reading across your email, files and tools. Before connecting an agent to sensitive systems, it is worth knowing what data it can access, what is retained, and who at the vendor could see it. Convenience and exposure move together here.
Bottom line
ChatGPT Work is a significant marker in the move from AI that talks to AI that does. Whether it genuinely "does your job" or becomes a powerful assistant you supervise depends heavily on the task and on how much unattended automation you are willing to trust. For now, treat it as a fast, capable junior worker: excellent at drafting and scaffolding, still in need of a review before anything it produces goes out the door.
This article summarises publicly reported details of OpenAI's ChatGPT Work launch (Bloomberg, Forbes and others, 9 July 2026). Product names, model variants and rollout details reflect that reporting; features and availability may change.
