Deep research moved into ChatGPT Work and Codex, searching the web, your uploaded files and connected apps, and returning an editable document with citations.
Account research has been a two-window job: deep research in one tab against the public web, and your own files, notes and connected systems somewhere else, stitched together by hand. Running it inside Work collapses that into a single pass, and the output arrives as an editable document with citations rather than a chat answer you have to reformat before anyone else will read it. The citation trail is the part that earns its keep on a first call, because a claim about a prospect you can source is a claim you can say out loud.
Deep research is the mode where the model goes away for several minutes, reads a great deal, and comes back with a long written answer instead of replying instantly. Until now it worked against the public web from an ordinary chat.
ChatGPT Work is the surface that can see your uploaded files and the apps you have connected to your account. Codex is the coding surface. Putting deep research inside Work means that same slow, thorough mode can read your material alongside the web in one run.
"Uses your existing Work or Codex allowance or credits" means there is nothing new to buy. It spends the budget your plan already gives you, and the separate deep research allowance in ordinary chat is untouched.
The short version: the thorough research mode can now see your files and connected apps rather than only the internet, and it hands back a citable document you can edit.
OpenAI made deep research available inside ChatGPT Work and Codex on September 9, 2026. In OpenAI's words, it can "research complex questions across the web, your files, and supported connected apps, then turn the findings into an editable document with citations."
Announced: September 9, 2026, in the ChatGPT release notes.
Availability: Plus, Pro, Business, Enterprise and Edu users with Work access, on web, desktop, iOS and Android.
Price: No separate charge stated. Deep research in Work and Codex draws on your existing Work or Codex allowance or credits, and OpenAI states that deep research limits in ordinary chat are unchanged.
Regions: Not stated.
What it doesn't do: Free and Go are absent from the eligibility list. OpenAI names no rollout schedule, no regional detail, and no figure for how many deep research runs a Work allowance covers, so the practical ceiling is whatever your allowance already was. The entry says "supported connected apps" without enumerating which ones, so no seller should assume a given CRM is among them without checking.
SDR — The gain is per-account speed on a named list. Point it at ten target accounts with your ICP notes attached and ask for one page each. The discipline to hold is the citation column: a personalised opener built on a sourced fact survives being forwarded to the buying committee, and one built on a plausible guess does not.
AE — This is your pre-call hour. Feed it the last three call notes and the account's public filings in the same run and ask what changed. The editable-document output is the real shift for you, because the artifact you need is one your manager and your SE can both open and mark up, not a chat you screenshot.
Manager — Decide what a good brief contains before the team runs a hundred of them. The output is only as consistent as the instruction, and a team producing a hundred differently-shaped research documents is worse off than one producing ten identical ones. Standardise the format first, then raise the volume.
RevOps — Two things to settle. Deep research in Work spends the Work allowance, so heavy research use now competes with everything else drawing on that budget — watch which teams burn it. And "supported connected apps" is unenumerated in the release note, so establish which of your systems are actually reachable, because the ones that are not are the ones reps will paste into a chat instead.
Research quality is set by the instruction, not the model: the difference between a page of facts and a brief someone acts on is whether you asked for a point of view and a source table. Start from a prompt that already separates fact, inference and seller hypothesis.
Two minutes, once a week. What changed in AI, and what to run because of it.