ChatGPT Work

ChatGPT Work is OpenAI's office agent inside ChatGPT, launched July 9, 2026 and powered by the GPT-5.6 frontier model with Codex technology built in. Unlike a chat assistant that answers questions, ChatGPT Work turns a goal into finished work: it gathers context across your connected apps, breaks a complex project into steps, and completes them independently, staying with a task for hours. Its output is finished material rather than conversation, including spreadsheets, slides, documents, and interactive web apps published as Sites. It serves business teams across sales, marketing, finance, operations, data, and engineering rather than any single vertical, and OpenAI reports nearly all of its own internal teams now use ChatGPT Work and Codex. Integration depth is broad. ChatGPT Work connects to apps through plugins, including Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRMs, and project trackers, and you can direct it to a specific source by typing the @ symbol and the app name. A unified plugins directory brings connectors into one place. On the desktop app it goes further: a built-in browser pulls in live web context, and Computer Use lets it operate your machine directly, clicking, typing, and moving files across your local apps and tools. It uses the Model Context Protocol as a client to reach external systems, though OpenAI does not publish its own MCP server. Codex, the coding agent used by more than 5 million people weekly, is merged into the same desktop app, so Chat, Work, and Codex share one surface. Pricing is the part OpenAI does not lead with. The minimum to use ChatGPT Work is free: it is available on every ChatGPT plan, including Free, through the desktop app for Mac and Windows. On web and mobile it rolled out first to Pro, Enterprise, and Edu on July 9, 2026, then to Plus at $20 per month and Business over the following days. The catch is that usage is metered on the same structure as Codex: a run draws on your plan's included usage, and more complex, longer tasks consume more of it. Business and Enterprise buy credits for usage beyond the included pool, and admins set spend controls in the Admin Console. So the real cost is not the flat plan price but the plan plus whatever agent usage the work requires, which is hard to predict for a tool designed to run for hours. What the free tier does not buy is meaningful sustained autonomy. Free and lower plans include only limited usage, so multi-step projects that run for hours will exhaust the included allowance and require a paid plan or purchased credits. The web and mobile experience also lagged the desktop launch, reaching Plus and Business only in the days after July 9. Enterprise pricing is custom and requires a sales conversation, and the credit-consumption model makes total cost modeling difficult, the same opacity that affects Codex-based billing across OpenAI's agent products. ChatGPT Work is not the best fit for teams that want a fixed, predictable per-seat bill, since usage-metered agent runs make monthly costs variable. Anthropic's Claude Cowork and Microsoft's Copilot and Scout compete directly in the office-agent category, and teams already standardized on Google Workspace or Microsoft 365 may prefer the agent native to that stack. For pure web-task automation rather than deliverable generation, OpenAI's own ChatGPT Agent, plus Browse AI, Skyvern, and Browser Use, target that narrower job. Current state Q3 2026: ChatGPT Work launched July 9, 2026 on web, mobile, and a new unified desktop app that merges the former Codex app; the previous desktop app is renamed ChatGPT Classic. Sites launched in public beta, and OpenAI began sunsetting its standalone Atlas browser in favor of a built-in browser and a ChatGPT Chrome sidebar extension. Named early adopters include Zapier, RingCentral, Virgin Atlantic, and NVIDIA. Security is built on the ChatGPT Enterprise foundation, with SOC 2 Type II, ISO/IEC 27001:2022, and ISO/IEC 42001:2023 confirmed on OpenAI's security page, an Auto-review layer that checks sensitive actions before they run, prompt-injection safeguards, and a Compliance API for admin oversight. Business and Enterprise tiers do not train on customer data.