OpenAI said on Thursday it is launching ChatGPT Work, based on GPT-5.6 and Codex technology, to carry out complex tasks.

ChatGPT Work gathers information across multiple apps and files to produce finished outputs such as documents, spreadsheets, presentations and web apps. With plugins, a built-in browser, computer-use and scheduled tasks, it handles repetitive work and long-running projects even while users are away.

ChatGPT Work runs on GPT-5.6, the latest model officially released on Thursday.

The scheduled tasks feature automatically performs repetitive work on a set schedule or when specific events occur. It can monitor customer feedback daily and organise it into prioritised product ideas, or automatically update presentations when new feedback arrives.

Plugins connect ChatGPT to Slack, Microsoft Teams, Google Drive, Microsoft SharePoint, email, calendars, customer relationship management (CRM) and project tracking tools, among others. Users can type '@' and an app name in a prompt to specify that context should be pulled from a specific app.

A Sites feature will also be released in beta. It turns work or ideas into interactive sites or web apps and shares them via a URL. It can be used for real-time dashboards, project tracking tools, prototypes and internal portals.

The desktop app will add a built-in browser and a computer-use feature. The computer-use feature performs actions such as clicking on-screen, entering content and moving files on the user's behalf.

OpenAI will also revamp its Chrome extension. It will gradually end service for Atlas, its existing ChatGPT-only browser. The existing desktop app will be renamed ChatGPT Classic. Existing Codex app users will be automatically switched to the new ChatGPT Work after updating.

On technical metrics, OpenAI said Sol, one of three GPT-5.6 family models, posted the best performance in coding, knowledge work, cybersecurity and science. It said it performs tasks with fewer tokens and at lower cost than previous and competing models.

In the "agent's final exam", which evaluates the ability to perform long professional tasks, Sol scored 53.6 points, 13.1 points ahead of Claude Fable5. Under a mid-level reasoning setting, it delivered performance 11.4 points higher at about one quarter of the cost.

In the "ArtificialAnalysis Coding Agent Index", it scored 80 points, the highest ever. Output tokens and task time were less than half of Fable5, and estimated cost was about one third lower.

In knowledge work evaluations, it set highs of 92.2 percent on BrowseComp and 62.6 percent on OSWorld 2.0. On OSWorld, it delivered higher performance while using 85 percent fewer tokens than Opus4.8.

In cybersecurity evaluation ExploitBench 2, it scored 73.5 percent, above GPT-5.5's 47.9 percent. In ExploitGym3, which measures the ability to implement real software vulnerabilities into working attack code, the 2-hour pass rate rose to 24.9 percent from GPT-5.5's 15.1 percent. On a 6-hour basis, it climbed to 33.7 percent.

ChatGPT Work will be rolled out from Thursday on web and mobile to Pro, Enterprise and Edu users in phases. It will expand to Plus and Business users within a few days. The desktop app will be available from Thursday to global Mac and Windows users, including a free plan.

Kyunghoon Kim (김경훈), head of OpenAI Korea, said, "ChatGPT Work shows that AI is evolving beyond a tool that answers questions into a partner that understands the user's context and completes real work together," and added, "It will be a first step to support companies and individuals in Korea to break free from repetitive work, focus on important problems and turn ideas into results."