From Chat Box to AI Productivity Colleague
ChatGPT Work is an enterprise AI agent that performs autonomous task automation across connected apps, web services, and local files, staying with a multi-step workflow automation project for hours as an AI productivity colleague rather than a short-answer chatbot. OpenAI has launched the ChatGPT Work agent as a new mode inside ChatGPT that can research, build documents, manage spreadsheets, and create web applications with minimal supervision. It is powered by GPT-5.6, the latest model family released on the same day after a limited preview period. In OpenAI’s own framing, this is the moment ChatGPT stops being a place to ask quick questions and starts being “a partner for your most ambitious work”. If you think this is just another UI mode, you’re missing the bigger shift: Work turns ChatGPT into something that behaves much closer to a digital employee than a chatbox.
What the ChatGPT Work Agent Actually Automates
The ChatGPT Work agent’s headline promise is autonomous task automation across the systems where knowledge workers live. It is designed for multi-step tasks that would normally require a human assistant: monitoring connected apps, running scheduled jobs, and producing finished deliverables. Users give a goal—such as analysing a monthly budget, preparing for a sales meeting, or turning research into a campaign brief—and the agent divides the work into smaller tasks, gathers information from connected applications, online services and local files, then outputs spreadsheets, presentations, documents, or web apps. This is multi-step workflow automation in practice: a single request can kick off customer research, flow into a campaign brief, and continue into market-specific materials while retaining earlier context. Crucially, the agent can remain active on longer projects for hours, researching and building applications independently while users drop in to review progress, answer questions, or approve key actions.
A Unified Agent That Lives in Your Stack
The most important detail for enterprises is not that ChatGPT Work is powerful, but that it is embedded. It connects directly to Slack, Microsoft Teams, Gmail, Google Drive, SharePoint, calendars, CRMs and project trackers, pulling workplace information without users copy‑pasting context. Typing “@” plus an app name tells the agent where to look, and a unified plugin directory centralises these connections. Scheduled Tasks allow recurring administrative work—reviewing new messages, refreshing meeting documents, checking dashboards, updating presentations on new feedback—to be handed off to the agent. Work follows people across web, mobile, and desktop so tasks can start on a phone, be checked between meetings, and resume on a laptop, with direct access to local files and installed apps plus a built‑in browser. This is what an AI productivity colleague looks like: persistent, integrated, and always plugged into your systems, not locked behind a browser tab.
Codex Inside: From Chatbot to Application Builder
Under the hood, ChatGPT Work is less a chatbot and more a control system for tools. OpenAI is merging its Codex developer tool directly into the desktop app for macOS and Windows, combining Chat, Work, and Codex in one interface. Codex technology is built into the service, so the agent can write code, use tools, and coordinate actions as a project develops. That matters because it turns the agent into a lightweight software factory: Work can not only draft a report but also assemble interactive web apps or internal tools, then publish them through the new Sites beta. One quotable benchmark highlights the ambition: “On Agents’ Last Exam, a benchmark covering long-running professional workflows across 55 fields, Sol scores 53.6, ahead of rival model Claude Fable 5 by 13.1 points”. In other words, GPT‑5.6 Sol is explicitly tuned to stay competent over long, complex workflows instead of one-shot answers.
Enterprise AI Agents: Power, Risk, and What Comes Next
ChatGPT Work represents a clear pivot from consumer chatbot to enterprise AI agents. OpenAI is positioning the service for work that spans several sources, tools and stages, not quick, single responses. More than five million people already use Codex weekly, with over one million using it beyond software development—a signal OpenAI interprets as demand for an agent that does the work, not only writes code. The GPT‑5.6 Sol, Terra and Luna models are rolling out over 24 hours, at published API rates of USD 5, 2.50 and 1 per million input tokens respectively (approx. RM23.00, RM11.50 and RM4.60), with corresponding output prices, and GPT‑5.4 will be retired on July 23 while GPT‑5.5 remains available. Access to ChatGPT Work lands first on web and mobile for Pro, Enterprise and Edu users, with Plus and Business following in the next few days. According to one source, the launch was even delayed by the Trump administration, highlighting how politically sensitive such agents have become. The broader trend is obvious: with competitors already moving from assistants to agents, ChatGPT Work is less a novelty and more the new baseline. The hard problem from here is governance—deciding how much of your workflow you let an AI productivity colleague run without you watching.
