OpenAI introduces ChatGPT Work, an AI agent that completes tasks autonomously

OpenAI has released ChatGPT Work, an AI agent that operates within the existing ChatGPT environment and can independently carry out multi-step projects. Rather than simply answering questions, the agent retrieves information from various apps and workflows, breaks an assignment into steps and delivers a finished result, such as a spreadsheet, presentation, document or shareable web application.

The launch comes alongside two other changes: the Codex app, previously a standalone tool for software developers, is being merged into the ChatGPT desktop app, and OpenAI is simultaneously making the GPT-5.6 model family generally available. ChatGPT Work is built on those new models.

Specific details about the exact release date and pricing had not been fully confirmed in the source material at the time of writing.

What ChatGPT Work does in practice

ChatGPT Work gathers context from workplace tools and then executes a series of steps without requiring the user to approve each action. The agent can connect to services such as Slack, Microsoft Teams, Google Drive and SharePoint via plugins. OpenAI references a directory of more than 1,400 plugins accessible through a single interface.

Through the desktop app, the agent also has access to local files and applications. Tasks can take hours to complete and are finished, including via scheduled assignments, without continuous human input. The output is a finished document, a presentation or a web app that can be shared immediately.

The integration of Codex into the desktop app makes code functionality accessible to a broader group of users. Professional developers will need to assess whether this works as well as a dedicated development environment, but for users without a programming background it lowers the barrier to generating software output.

The GPT-5.6 model family behind the agent

ChatGPT Work runs on the GPT-5.6 family, which OpenAI released at the same time. The family consists of three variants: Sol is the most capable, Terra is aimed at balancing performance and cost, and Luna is the cost-efficient option. According to OpenAI, all three are optimised for multi-step tasks, reasoning and the production of deliverables.

The model selection determines which variant the agent deploys for a given task. For those building with AI, this is relevant: the availability of three tiers within a single integrated environment makes it easier to weigh cost against capability on a per-task basis, without switching platforms.

Competition with Anthropic and the broader agentic market

ChatGPT Work is entering a market in which several major providers are developing comparable agent functionality. Anthropic released Claude Cowork earlier this year, based on a similar principle. Technology companies are currently investing heavily in AI systems capable of partially taking over work processes with minimal user intervention.

With this move, OpenAI is positioning ChatGPT as a broader work environment rather than merely a chat window. Whether that positioning resonates with companies and organisations that have their own security requirements and existing tool stacks will depend on factors such as data governance, compliance and integration capabilities, areas that had not been fully elaborated in the available source material.

What this means for those working or building with AI

For founders and developers using AI in their products or processes, the bar shifts with every iteration of these kinds of agent systems. ChatGPT Work demonstrates that the endpoint of an AI interaction is no longer a block of text, but a usable file or a working application.

At the same time, the integration of so much functionality into a single platform raises questions about dependency on a single provider. Companies that now integrate deeply with OpenAI's plugin directory are building a connection to a platform that is changing shape at a rapid pace. The merger of Codex into the main app is itself an example of this: what was a separate product yesterday is today a component of a broader whole.

Those involved in the launch include product manager Ty Geri, Codex lead Andrew Ambrosino and OpenAI engineer Thibault Sottiaux, alongside CEO Sam Altman and president Greg Brockman.