ChatGPT Work brings long-running tasks into ChatGPT

For years, ChatGPT has been an interface for asking questions, summarizing material, or drafting a first version of a text. OpenAI now wants it to be a place where a long assignment can remain active: gathering context, organizing a plan, working with connected files and apps, and returning a result that can be reviewed. That is the idea behind ChatGPT Work, introduced by the company on July 9.

The proposal is not simply a longer answer in a chat. Work is designed for multi-step tasks and deliverables such as documents, spreadsheets, presentations, reports, and Sites. OpenAI says it can stay with a project for hours, break it into parts, and advance independently. The aim is to bring agent behavior to everyday work that does not begin and end with one question.

From quick chat to work that retains context

The main difference is the kind of assignment. For a short question, search, or brainstorm, OpenAI keeps Chat as the fast conversational experience. Work is intended for researching a topic, analyzing information, and producing finished material. A project can hold related conversations, files, and instructions, so a new Work thread can begin from that context rather than asking for it again from the start.

In its announcement, OpenAI says Work can gather information from connected apps and workflows. Its practical value depends on the quality of the context. A good assignment still needs to explain the goal, constraints, expected format, and review criteria. An agent does not turn an ambiguous request into a well-defined business decision, but it can reduce the mechanical work of gathering, organizing, and transforming available material.

The new mode also includes scheduled tasks. According to the release notes, a project can run once, repeat on a schedule, or monitor for changes. That opens uses such as updating a recurring report or following an information source. The important point is not only that a task runs on its own. The user can review progress, change direction, and approve actions that require approval.

Work does not replace Codex

The launch comes with a distinction that OpenAI states clearly. Codex remains the dedicated experience for software development: it works with repositories, local folders, terminals, tests, and developer tools. Work, by contrast, is aimed at research, analysis, and materials that can involve documents, spreadsheets, or presentations.

That separation helps people choose a tool by the desired outcome rather than by a fashionable label. If the goal is to fix code, run tests, or review repository changes, Codex remains the natural place. If the goal is to turn a collection of sources and files into a report or proposal, Work is the experience OpenAI proposes. Both can handle complex tasks, but their interfaces, context, and controls are oriented toward different kinds of work.

OpenAI links Work to Codex technology and GPT-5.6, but that does not mean every task runs in the same way or that access is identical on every surface. The official help center says Codex retains a separate view and history in the desktop app. That distinction matters for teams moving between business analysis and software development, because clear boundaries reduce the risk of using more permissions or more context than a task needs.

Files, apps, and devices: permission is part of the product

Work runs on the web and mobile as a cloud experience for eligible accounts, and it is also available in the desktop app when a plan and workspace include it. Cloud Work conversations can continue across web, mobile, and desktop. Local conversations, by contrast, stay on the computer.

The distinction is significant. On desktop, Work can use local files and desktop apps with user permission. That makes privacy a concrete decision for every task. It is not enough to think about what should be produced; people also need to decide which folders, documents, or services the assignment actually needs. A sensible approach is least-privilege access rather than enabling connections for convenience.

OpenAI safety guidance warns that a connected agent can encounter sensitive data and act on a user behalf. It recommends enabling only necessary apps, reviewing permissions, and avoiding open-ended requests such as handling all email. It also notes that some actions and access requests need human intervention or confirmation. A well-configured workflow does not remove human judgment. It focuses it on the moments with the greatest impact.

What changes for an individual or team

The most interesting promise of Work is not that it can write a document. Many tools can already do that. It can preserve continuity between research, drafting, review, and update, within a project that keeps its materials and rules together. For an individual, that can reduce repeated instructions. For a team, it can make the criteria that turn scattered notes into a useful deliverable more visible.

There are practical limits. Availability is rolling out gradually to eligible accounts and varies by plan, surface, and workspace configuration. Result quality depends on the sources, enabled apps, and human review. An agent can accelerate a task, but it should not receive open-ended authority or indiscriminate access to private data.

Autonomy that remains reviewable

ChatGPT Work shifts the focus from a conversation that answers to work that persists and produces. The approach can be valuable when an assignment combines scattered information and needs several iterations. Its best use is not asking it to do everything without context, but giving it a bounded goal, relevant material, and a clear review point.

That balance is the most useful part of the launch. Automation can free time from coordination and preparation. Responsibility for deciding what information is shared, which action is approved, and when an output is ready remains human.