What Is ChatGPT Work? How It Differs from Chat, Codex, Classic, and the API

ChatGPT Work is an agent inside ChatGPT, not a new standalone AI tool. ChatGPT remains OpenAI's general assistant; Work is its outcome-oriented lane for longer research, analysis, workflows, files, and finished deliverables. Codex remains the dedicated coding agent, while regular Chat remains the fastest place for questions and conversation.

The useful boundary is the job being delegated. Chat helps you think through a problem turn by turn. Work accepts a broader objective, plans the work, uses tools and sources, and returns something reviewable. Codex applies the same agentic foundation to software repositories, terminals, tests, and technical delivery. These surfaces can cooperate, but they are not interchangeable names for one product.

What ChatGPT Work is

Work is designed for assignments that take more than one response or one source. You give it an outcome, relevant context, and constraints; it can break the assignment into steps, research across sources, use connected tools, and assemble a document, spreadsheet, presentation, site, or other deliverable. You can inspect progress, redirect the approach, and review the result rather than manually coordinating every intermediate prompt.

That makes Work a fit for jobs such as synthesizing research into a briefing, reconciling information across files and services, preparing a decision memo, building a recurring reporting workflow, or turning a rough objective into a finished set of materials. It is not automatically the best choice for every question. If the value comes from a quick answer, a short draft, brainstorming, or conversational search, regular Chat has less setup and keeps the interaction immediate.

OpenAI says Work is built with the same core agent technology as Codex. That describes the execution foundation, not the product boundary. Work is tuned for ambitious everyday knowledge work across sources and applications. Codex is tuned for engineering work where the central objects are code, repositories, development environments, tests, and deployments.

Who gets Work and where it runs

Work is rolling out by surface and plan rather than appearing everywhere at once. OpenAI's launch sequence puts web and mobile access first on Pro, Enterprise, and Edu, with Plus and Business following. The release notes also refer to Pro Lite in the early rollout. Managed workspaces can add preview, administrator, and policy controls, so an eligible plan does not guarantee that every member sees Work on the same day.

The new desktop app has the broadest plan reach. OpenAI says it is available globally on macOS and Windows across ChatGPT plans, including Free, while the live pricing page describes Free and Go access as limited. Plan, account, workspace, region, platform, and rollout state can therefore change what a particular user sees.

This is why Work should not be described as a universal entitlement attached to the word ChatGPT. The maintained ChatGPT pricing page and in-product account controls are the right sources for the current plan matrix. An evergreen explainer can define the surface and buying boundaries, but it should not freeze a fast-changing rollout into a permanent promise.

The new desktop app and ChatGPT Classic

The new ChatGPT desktop app brings Chat, Work, and Codex into one application. It is an evolution of the Codex desktop app, with the general ChatGPT experience added alongside the two agent surfaces. Work on desktop can use local files, approved applications, a built-in browser, and other computer tools, which makes it better suited to tasks whose context lives on the machine rather than entirely in the cloud.

The previous ChatGPT desktop application is now called ChatGPT Classic. Classic is a legacy client path, not a fourth kind of agent and not a higher or lower subscription tier. OpenAI says it will continue to receive support and updates, but new agent features are centered in the new desktop app. Choose Classic only when the familiar client or a compatibility need matters more than Work and Codex integration.

Cloud and desktop execution also have a practical separation. Work on web or mobile runs in the cloud, while desktop Work can act against local context and applications. OpenAI says cloud and desktop Work threads do not initially sync. If a task must move between devices, keep the source files and handoff notes explicit instead of assuming the same agent thread will follow automatically.

Plugins, browser, and computer use

Plugins extend what Work can do without turning every integration into a new canonical product. OpenAI plugins can package instructions and skills, connected apps or MCP tools, hooks, browser extensions, and scheduled-task templates. They are available in Work on the web and in Work and Codex on desktop, subject to plan, role, workspace policy, and the permissions of each connected service.

A plugin does not bypass authorization. Access to email, documents, calendars, databases, or internal services still depends on the user's account and the organization's controls. A well-scoped plugin should expose only the tools and data needed for the job. Users should inspect requested permissions and keep approval checkpoints around publishing, sending, purchasing, deleting, or changing external systems.

Work can browse in more than one way. Cloud Work can use a hosted browser for public web research, but OpenAI's launch documentation limits that route to public sites and says it cannot sign in or complete payments. The desktop app includes a built-in browser with its own profile and state. A Chrome extension can instead let the agent work with an existing Chrome profile when the user explicitly enables it. These routes differ in session state and risk; “has a browser” does not mean unrestricted access to every logged-in site.

Computer Use is the desktop capability for seeing, clicking, and typing across approved applications. OpenAI documents it for Work and Codex on macOS and Windows in supported regions, with a plugin and operating-system permissions required. It is useful when a workflow crosses apps that lack a direct connector, but direct tools or APIs are usually more reliable for structured operations. Sensitive actions should remain visible, narrowly authorized, and reviewable.

Usage, models, and the Ultra boundary

Work and Codex share the agentic usage and credit structure attached to an eligible ChatGPT account. Shared does not mean one Work task equals one Codex message, or that published Codex examples predict the cost of a Work assignment. Consumption varies with the selected model, input and output tokens, context size, reasoning effort, tools, task complexity, and how long the agent runs.

The model boundary also differs by plan and surface. OpenAI's GPT-5.6 announcement says Free and Go receive Terra in Work and Codex, while Plus, Pro, Business, and Enterprise can choose among Sol, Terra, and Luna. Ultra is deliberately narrower: Work Ultra is listed for Pro and Enterprise, while Codex Ultra is listed for Plus and higher. A user should never infer Work Ultra access from Codex access, or assume every ChatGPT plan receives Ultra.

Exact allowances, credit options, and prices belong on the maintained ChatGPT and Codex pricing pages. They change faster than the concept of Work, and a full copied plan table would blur rather than clarify the decision. Check both pages when estimating a heavy workflow because one shared pool can be consumed by activity in either Work or Codex, even though individual tasks do not use a fixed interchangeable unit.

The OpenAI API remains a separate commercial route. A ChatGPT subscription does not include API usage, and API billing is managed separately from ChatGPT billing. Signing into Codex with an API key can put eligible local Codex work on usage-based API billing, but that does not turn the API into a Work entitlement and does not make ChatGPT subscription limits available to an API integration.

Work, Chat, Codex, Classic, or the API

Choose Work when the desired output is more important than maintaining a turn-by-turn conversation and the task needs several coordinated steps. Choose Codex directly when code is the primary working material or repository-aware engineering controls matter. Choose regular Chat when you want to think with the assistant rather than delegate a project. Choose Classic only for the legacy desktop experience. Choose the API when you are building a product or automation, not merely asking ChatGPT to complete work on your behalf.

A practical decision rule

Start with the smallest surface that matches the job. Use Chat to clarify the objective or test a question. Move to Work when the assignment needs planning, multiple sources, tools, a durable deliverable, or scheduled repetition. Move to Codex when the definition of done includes code changes, tests, repository state, or a technical environment. Move to the API when the workflow must run programmatically inside your own system.

For Work, provide the outcome, audience, allowed sources, output format, constraints, and approval points. Prefer direct plugins or APIs for structured systems, use browser or Computer Use when no better interface exists, and require confirmation before consequential external actions. Review citations, calculations, files, and side effects before treating the result as finished.

Finally, check the live pricing and rollout pages before committing a team workflow. The important questions are not only “Do I have Work?” but also “Which surface will execute it, which model and reasoning level can this plan use, which permissions are enabled, which shared agentic pool will it consume, and does this job belong in ChatGPT or in a separately billed API system?”