On July 9, 2026, OpenAI announced "ChatGPT Work," an agent that spans multiple business apps and local files to produce finished deliverables. The new ChatGPT desktop app now lists Chat, Work, and Codex side by side, available across all plans on Mac and Windows starting the same day. The previous standalone Codex app is being folded into this, and the existing ChatGPT desktop app will be renamed "ChatGPT Classic." Additionally, the standalone Atlas browser is heading toward discontinuation. This announcement is less about adding a feature and more about OpenAI's product reorganization to consolidate the entry point for work software into ChatGPT.
A Desktop Reorganization That Folds In Codex and Atlas
ChatGPT Work receives a goal, gathers the necessary information, breaks the work into smaller steps, and continues processing tasks that take hours. OpenAI explicitly states "Codex technology built-in." This design extends the agent execution infrastructure that began with Codex for developers to a broad range of job functions, including finance and sales. According to the company, Codex now has over 5 million weekly users, with more than 1 million of them using it for work outside of software development.
This integration also changes product boundaries. The desktop app handles local files and apps, while the built-in browser opens websites and cloud-based documents. Computer Use handles clicks, input, and file movement on the user's behalf. While out and about, users can start, check, and course-correct tasks from mobile. In effect, the separate pathways of Codex, Atlas, and the previous version of ChatGPT can now be moved between within a single task.
While availability is broad, the computing resources accessible are not uniform. The desktop version is available to all plans, including Free, while Web and mobile start with Pro, Enterprise, and Edu, expanding to Plus and Business within days. Usage follows the same mechanism as Codex, meaning longer, more complex tasks can consume a larger portion of a plan's allotment. Rather than adding a new standalone subscription product, it's more accurate to say a heavy-duty task execution layer has been added within existing plans.
Over 1,400 Plugins Connect Work Through to Deliverables
ChatGPT Work is not a feature for answering questions—it's a mechanism that connects the entire flow from scattered information to finished deliverables. The integrated plugin directory offers more than 1,400 plugins. These connect to communication tools like Slack and Microsoft Teams, file-sharing services like Google Drive and SharePoint, and CRM and project management tools. ChatGPT determines which plugins are needed based on the prompt, and users can also specify a reference target using "@".
Output destinations have also expanded beyond the chat window. In addition to documents, slides, and spreadsheets, the publicly beta Sites feature lets users share dashboards and progress-tracking screens as web apps. These can also be used for internal portals or interactive reports. If the underlying data changes, updates can be handled automatically. With Scheduled Tasks, users can set recurring processes—such as revising meeting materials based on Slack updates, reporting dashboard changes every morning, or updating product proposals based on customer feedback—triggered by time or events.
However, these components didn't appear out of nowhere. In July 2025, ChatGPT agent could use browsers and terminals to compile information obtained from connectors into slides and spreadsheets, and it also supported scheduled execution. That October's Company Knowledge could span Slack, SharePoint, and other services, but enabling it disabled web search and image/graph generation. In spring 2026, Workspace Agents and sidebars for Excel and Google Sheets were added, followed by write access to apps and admin-facing analytics. What's new with ChatGPT Work is that it reduces the burden of finding and switching between these tools, allowing research, creation, and updates to persist as part of the same ongoing task.
GPT-5.6 Elevates the Quality of Deliverables
A work agent that finds the right information but produces sloppy spreadsheet formulas or poorly structured slides doesn't actually reduce the time humans spend on corrections. GPT-5.6 is responsible for this finishing touch. OpenAI explains that it transforms miscellaneous context from documents and daily tasks into shareable deliverables, and can review and correct generated screens using Computer Use.
According to figures published by the company, GPT-5.6 Sol scored 62.6% on OSWorld 2.0, which measures desktop operation. On BrowseComp, which measures difficult web research tasks, Sol scored 90.4%, while Sol Ultra, which parallelizes multiple tasks, scored 92.2%. Since the 2025 ChatGPT agent scored 68.9% on BrowseComp, this represents a significant improvement, at least in research evaluation. However, these are OpenAI's own evaluations, and they don't reveal how many consecutive hours-long internal business tasks the system can complete, or how many minutes corrections take.
Model options also vary by plan. Free and Go use Terra, while Plus and above can choose between Sol, Terra, and Luna. The "ultra" tier of ChatGPT Work is available to Pro and Enterprise users. Even though Work is a single entry point, speed and reasoning capacity vary by contract, and the available capacity for parallel execution also differs. When piloting the feature, what matters isn't the presence or absence of functionality, but measuring the actual usage consumed to complete real tasks and the human correction effort required.
Expanding Permissions Also Expands the Attack Surface
Connecting to over 1,400 services increases convenience, but it also demands a stricter design to prevent misoperation and data leaks. ChatGPT's app permissions can be set to confirm every action, to automate reads while confirming changes, or to operate without any confirmation at all. The default "Important actions" setting automates reads while requiring approval for high-impact operations such as sending or deleting emails, making purchases, or changing sharing permissions. "Never ask" operates without any confirmation, and OpenAI itself labels this as carrying high risk.
Data handling also varies by plan. For Business, Enterprise, and Edu, information retrieved from apps is not used for model training by default. For the four individual plans (Free/Plus/Go/Pro), data may be used for training if "Improve the model for everyone" is enabled. Since Work is offered across all plans, individual users should check this setting and each app's terms of service before connecting. On the enterprise side, role-based connection controls, action confirmation, and audit logs are necessary.
OpenAI announced that Auto-review, which requires prior confirmation for critical actions, blocked 100% of attacks attempting to extract protected data in adversarial red-team testing. However, it should be noted that this is an internal evaluation with a limited testing scope. In its 2025 ChatGPT agent announcement, the company explicitly noted the danger of prompt injection, where hidden instructions on a webpage could cause connector data to be sent externally. In 2026, it also introduced Lockdown Mode, which restricts live web access, Agent Mode, and live connectors. A operational trade-off between connectivity and safety still remains.
The Significance of Competing With the Same Model as Microsoft 365
GPT-5.6 is not exclusive to ChatGPT Work. It's also the priority model for Chat and Cowork in Microsoft 365 Copilot, in addition to Word, Excel, and PowerPoint. If both products use the same model lineage, what enterprises should compare is not the model name, but where the work context is drawn from, how far the system can act, and who can oversee that process.
OpenAI has taken a horizontal approach, spanning different SaaS products and local PCs while letting users select the tools they need from over 1,400 plugins. Microsoft's Work IQ, by contrast, assembles context from Microsoft 365 and external systems in accordance with tenant-level permissions. Through MCP, it consolidates hundreds of operations—including email, calendar, files, and chat—into 10 general-purpose tools, with every call logged and evaluated. For enterprises where data and business processes are concentrated in Microsoft 365, this integration is powerful. For enterprises spanning multiple SaaS products and local apps, the breadth of ChatGPT Work is effective.
What should determine adoption decisions isn't a flashy demo, but day-to-day reproducibility. Organizations should measure the completion rate for repeated tasks and the usage consumed per deliverable. Recording the number of times humans had to send work back for correction, along with the time needed to trace a misoperation through an audit, will reveal which solution truly shortens an organization's work. What OpenAI should demonstrate going forward is the success rate and cost of long-duration tasks, as well as operational data proving that its 1,400+ connections can be governed in a production environment.