OpenAI Launches ChatGPT Work: Autonomous AI Agent for Cross-App Tasks

OpenAI released ChatGPT Work, an agent that gathers data across apps, splits projects into steps and runs for hours on its own using GPT-5.6.

The update lands on July 10, 2026 and targets Pro, Enterprise and Edu customers first, with Plus and Business plans added in the following days. As OpenAI states on its blog, the agent “can collect information across apps and files, break complex projects into small steps, and keep working for hours on its own.”

Desktop ChatGPT now shows Chat, Work and Codex modes for all users, including free accounts, after Codex was folded into the main desktop app.

OpenAI states that Codex already reaches more than five million weekly users, with over one million of those uses outside software development.

The move puts pressure on other agent builders that still require users to manage each step or stay inside single applications.

Desktop mode availability for free users removes a previous barrier and lets more people test the full Work agent without paying.

OpenAI positions ChatGPT Work against tools that perform isolated tasks but cannot sustain long multi-step projects across separate programs.

The agent draws from Codex technology and the new GPT-5.6 model. GPT-5.6 improves multi-step reasoning by maintaining an internal state across extended sessions and switching context between applications without losing task history, enabling the agent to follow templates and maintain coherent reasoning across several hours. This approach tests whether a single company-controlled model can replace fragmented agent frameworks that combine separate tools.

Work mode can open files, collect information from different applications and produce draft materials without new prompts for each sub-task. A product manager at a mid-size consulting firm described using it to pull quarterly data from spreadsheets and CRM tools into a client deck: “I started the task, minimized the app, and returned to a completed first draft three hours later with sources cited.”

Users report that the agent continues running when the desktop app is minimized, returning completed sections later. A freelance researcher noted similar behavior when compiling literature across browser tabs and PDF readers without re-prompting.

The feature set matches the description shared by OpenAI on its site for ambitious work that spans documents, data and applications.

Some observers question whether GPT-5.6 can maintain accuracy over several hours when context grows or when source files contain conflicting data.

Independent tests will need to confirm if the agent updates its plan correctly when external applications return unexpected results.

OpenAI has not released detailed error rates for Work mode on mixed file types.

The next three months will show whether weekly active usage of Work mode grows inside Enterprise accounts and whether competitors match the multi-hour persistence.

Desktop metrics for free users will indicate if broader access leads to higher conversion or simply higher basic usage.

Any public benchmark from OpenAI or third parties on step-completion accuracy will clarify the agent’s practical limits.

Readers who handle repeated cross-application projects should test Work mode once it reaches their plan to measure time saved against current manual workflows.