Deep Research is now available in ChatGPT Work and Codex
Paid ChatGPT users can now run extended research sessions in Work and Codex, pulling from the web, files, and connected apps, with output as editable documents, presentations, spreadsheets, or Sites.
From 9 September 2026, Deep Research is available inside ChatGPT Work and Codex. If you are on a Plus, Pro, Business, Enterprise, or Edu plan with Work access, you can now run an extended research session that pulls from the public web, files you upload, and supported connected apps, and then produce the result as an editable document, presentation, spreadsheet, or Site.
That is a meaningful step forward from Deep Research’s original form. When OpenAI launched it in February 2025, it was a standalone report generator built on o3, useful but somewhat isolated. It produced a document you could read. Now it sits inside a broader working environment, alongside your files and connected tools, and the output is something you can directly edit and share.
What Work and Codex actually do
It helps to be clear on what these two modes are for. Work is ChatGPT’s environment for knowledge tasks: researching a topic, analysing information, and creating documents, spreadsheets, presentations, reports, or Sites. Codex handles technical work: writing and debugging code, running tests, reviewing changes, and working with repositories.
Deep Research is now available in both. In Work, that means you can ask a complex question, let the agent gather and cross-reference evidence from multiple sources, and receive a cited deliverable you can refine. In Codex, the same research capability sits alongside your technical work, which makes it useful for things like a migration brief, a dependency review, or an implementation decision you want to ground in real evidence before touching the codebase.
How to use it
To start a deep research task in Work or Codex, type @Deep Research in the chat window. On any supported client, you can also ask for deep research explicitly in plain language. The agent will ask clarifying questions before it begins, and you can steer it with additional instructions while it is running. When it is done, you specify the output format you want: document, presentation, spreadsheet, or Site. Which formats are available depends on the tools enabled for your workspace and the task at hand. Google Docs, Sheets, and Slides are available if you have the Google Workspace connection set up.
The whole flow, from initial prompt to cited deliverable, happens inside the same environment where your team is already working.
The credits question
One thing worth understanding before you start is how allowances work, because there are two separate meters involved.
When you run Deep Research in ordinary Chat, it draws from Chat’s own plan-dependent research-task allowance. When you run it in Work or Codex, it uses the Work/Codex allowance or credit pool your team already shares for agent tasks. A Work research task does not touch your Chat research allowance, and vice versa. The two pools are independent.
For Enterprise and Edu admins, Deep Research permissions are managed under Workspace settings, in Permissions and roles. Web search also needs to be enabled. As with all Work and Codex tasks, the feature does not grant any additional data access beyond what the connected provider, the user’s account, and workspace policy already permit.
Compliance logging differs by mode
This is a detail that matters if your organisation has compliance requirements. Deep Research run in Chat is covered by ChatGPT’s conversation logging in the OpenAI Compliance Platform. Deep Research run in Work or supported Codex clients uses Codex compliance logs, which include prompt and response text and supported app activity. If your team operates under specific audit or retention obligations, it is worth confirming which mode your researchers are using and whether the corresponding logging is configured correctly.
What this means for your team
The practical implication is that a significant chunk of knowledge work, gathering evidence, synthesising sources, and producing a shareable deliverable, can now happen inside the same tool your team uses for everything else. You are not context-switching between a research tool, a document editor, and a project workspace. The agent handles the path from scattered evidence to a reviewable output in one session.
For teams using Codex, the combination is particularly useful. A developer preparing a significant architectural change can ask Deep Research to pull together relevant documentation, prior decisions, and technical context, and have a brief ready to review before writing a line of code. The research and the technical work sit in the same place.
Deep Research has also been updated significantly since its original release. Since February 2026 it runs on a model based on GPT-5.2, rather than the original o3 base. A February 2026 update also added the ability to connect it to MCP sources and restrict web searches to trusted sites, so if your organisation needs research grounded in specific, authoritative sources rather than the open web, that is configurable.
The direction OpenAI is moving is fairly clear: rather than building a separate productivity suite to compete with Microsoft 365 or Google Workspace, they are making one agent capable of coordinating research, sourcing files, connecting to apps, and delivering a finished output. Whether that approach serves your team better than the tools you already have is a practical question worth testing with a real task rather than a theoretical one.