If you searched for OpenAI ChatGPT Agent, note that OpenAI’s current Help Center says ChatGPT Agent is no longer available and points users to ChatGPT Work for longer, multi-step tasks and finished deliverables. To try Work, open ChatGPT on web or mobile and start a task in Work. OpenAI specifically documents this route for cloud-browser tasks. Cloud-browser access is available on paid ChatGPT plans in supported regions, excluding Free and Go; availability can vary during rollout and depend on workspace permissions. Work’s available options can also depend on your plan and workspace settings, so check the current ChatGPT Work guidance and cloud-browser availability for your account. The Agent help page still contains older /agent and plan-availability instructions, so don’t rely on those legacy steps as current access guidance. (ChatGPT Agent Help Center)
ChatGPT Agent was a ChatGPT feature, not an API model name. If you want ChatGPT to do a task for you, investigate Work. If you’re building an application, choose a developer API path separately; ChatGPT’s interface and tools do not automatically carry over to an API.
ChatGPT Work or a production API?
| Your goal | Starting point | What to verify |
|---|---|---|
| Have ChatGPT complete and return work for you | ChatGPT Work | Open ChatGPT on web or mobile and start a task in Work. For cloud-browser tasks, availability is limited to paid plans in supported regions (excluding Free and Go), and may depend on rollout and workspace permissions. Other Work options can depend on your plan and workspace settings. |
| Build a long-running agent into an application | Agents API | It uses an OpenAI-managed agent harness; your application provides tools and chooses the execution environment. Check its current beta status, prerequisites, and limits. |
| Build a bounded file-analysis workflow with direct control over model requests | Responses API | Choose a model and tools documented for your file types, and manage the integration and any workflow state your application needs. |
OpenAI’s developer guide distinguishes these options by who runs the agent loop and manages state: the Agents API is for longer-running work with a managed harness; the Agents SDK runs in your application; and the Responses API gives you direct control over model calls or a custom agent loop. (Agents overview)
A documented starting point for file-to-report work
For a multi-step workflow—such as comparing several proposals and producing a report—one documented option to evaluate is the Agents API with the OpenAI-hosted sandbox. The current quickstart uses gpt-6-astra with that environment for a task that writes and runs a script. The Agents API overview says agents can work with files and execute code in a sandbox. That makes it a concrete starting configuration to test when your application can make the relevant files available to the chosen environment; it is not evidence that this model is best for every report or file type. (Agents API quickstart; Agents API overview)
If your task is primarily to analyze documents rather than coordinate long-running work, also evaluate the Responses API file-input path. OpenAI’s file-input guidance says PDF processing uses text and page images on vision-capable models, giving GPT-4o and later models as examples. It says non-PDF documents are processed as text, and suggests File Search for retrieving from large files. Check the current model documentation and file-format guidance for your actual inputs before selecting a model. (File inputs; File Search)
For the Agents API, the quickstart also specifies application API-key permissions for session operations and model inference. Keep the key in your application—not in the agent’s sandbox—and confirm the required permissions in the official quickstart. The documented gpt-6-astra example is a starting configuration, not a comparative model recommendation or a guarantee of file compatibility.
Test the workflow before production
Use representative inputs and check the delivered result, not just the agent’s final message. For a proposal-comparison workflow:
- State which files the agent may use, the report format, and actions it must not take.
- Test a normal case with complete proposals and an exception case with missing or conflicting information.
- Check that the report cites or identifies its supporting material and flags uncertainty rather than inventing a missing detail.
- If the application writes or shares the report, inspect the result in its intended destination and define which consequential actions require human approval.
These are practical validation steps, not claims that a particular API provides a guaranteed approval barrier or error-free result. Confirm the available tools, environment, and review controls in the documentation for the path you choose.
Bottom line
The old ChatGPT Agent feature is marked unavailable in OpenAI’s current Help Center; ChatGPT Work is the stated path for longer, multi-step tasks in ChatGPT. For an application, evaluate the Agents API when you need its managed long-running harness, or the Responses API for a more directly controlled file-analysis workflow. Choose a model only after matching its documented capabilities to your actual file types, then test ordinary and exception cases before relying on the output.
Sources
ChatGPT agent — OpenAI Help Center
ChatGPT Work and Codex — OpenAI Help Center
Using cloud browser in ChatGPT — OpenAI Help Center