ChatGPT deep research

Deep research is not a separate product with its own login. It is a mode of ChatGPT started by typing /Deepresearch, picking it from the + tools menu, or choosing it in the sidebar, as the help article sets out. The run is a sequence: you state the outcome, choose sources, review a proposed plan, watch the run and interrupt it, then get a report with citations.

Sources are the part worth attention. By default it reads the public web and files you upload. Through Sites > Manage sites you can restrict a run to listed domains, or select "Prioritize these sites, but allow full-web search". Connected apps such as Google Drive or SharePoint and authenticated data services can also be sources, and only their read actions are used: "It does not use app write actions as part of research."

Finished reports open full screen with a table of contents, a sources section and an activity history, and download as Markdown, Word or PDF.

Every ChatGPT plan includes deep research, and the pricing page grades it in words rather than numbers: "Limited deep research" on Free, "Expanded deep research" on Plus, "Maximum deep research" on Pro. Go, at €8 (about $8.95) a month, is not credited with deep research in its own card at all, while the comparison grid shows Limited for the two lowest columns.

The help article pushes the number into the product: "Deep research usage varies by plan. Your in-product usage counter shows your remaining tasks. For plans with a fixed monthly allowance, it resets every 30 days from the date of your first use." So a buyer cannot size the feature before subscribing, and the reset is anniversary based rather than calendar based.

A second FAQ answer notes availability itself is conditional: asked whether deep research works in all countries, the answer is "No. Availability depends on your plan and your country or territory."

Pricing is stated as "priced per user per month" in the FAQ. Read from this location the individual cards show €0 Free, €8 (about $8.95) Go, €23 (about $26) Plus and from €103 (about $115) Pro, all with a / month label; OpenAI serves currency by region, so a reader elsewhere will see local figures.

Business is the clearer case because the card prints both terms in text rather than behind a toggle: a standard seat is "€21/month (about $24) if billed annually. €26/month (about $29) if billed monthly", and a premium seat "€105/month (about $118) if billed annually. €130/month (about $146) if billed monthly" with "5x more usage than standard, with no 5-hour limit". Business starts at two users. Enterprise is "Custom pricing" through sales, with invoicing and volume discounts.

On top of the seat there is a consumption layer: the grid footnote reads "Enterprise and Business can purchase credits for more access", and the Enterprise card mentions credit-based and token-based options. Nonprofits are offered up to a 75% discount on Business or Enterprise.

For Enterprise and Edu workspaces, access to the feature runs through role-based access control, and the help article is careful about what enabling an app does and does not do: "Enabling an app for a workspace does not guarantee that it supports deep research or that every member can access the same provider content." Research also respects the permissions of the connected account, so it cannot read what the signed-in user cannot.

Auditing exists at the workspace level. Deep research activity is "detailed in the Conversation API, along with the conversation where the task was started", which is the Compliance API that the pricing grid lists as an Enterprise row.

Published assurances are SOC 2 Type 2 for business plans, ISO 27001, 27017, 27018 and 27701 for Enterprise, SAML SSO and MFA on Business, SCIM and Enterprise Key Management above that, and data residency in US, EU, UK, JP, CA, KR, SG, IN, AU and UAE.

"Conversations using deep research follow the same data handling and privacy settings as regular ChatGPT conversations." That cuts both ways. Consumers can switch off "Improve the model for everyone" and keep their history, and business data is not trained on by default. OpenAI will execute a DPA for Business, Enterprise and the API.

Retention is where the plans diverge. Deleted chats are scheduled for permanent deletion within 30 days and temporary chats may be kept up to 30 days for safety, but a configurable retention window is named only for Enterprise, Healthcare and Edu. Export diverges the same way: personal accounts can request their data, while "Self-service data export is not available for ChatGPT Business, Enterprise, or ChatGPT for Healthcare workspaces".

No on-premise option is mentioned anywhere read, and no license or source repository is named.

Programmatic deep research is not this feature. It is a set of models on the API platform, used through the Responses API with at least one data source: web search, remote MCP servers, or file search over vector stores. The same page carries a deprecation notice naming a shutdown date for o3-deep-research and o4-mini-deep-research with a general model as the replacement, and warns that "changing the model ID alone is not a complete migration".

So a team that wants deep research in its own software is buying an API model family in transition, not the ChatGPT surface described above. Background mode, recommended for these long runs, "retains response data for roughly 10 minutes so that polling works reliably, which makes it incompatible with Zero Data Retention (ZDR) requirements".

It fits analysts and professional teams who need a documented, citation-backed report over a mixed source set, and who care which sources were used: approved domains, uploaded files and connected document stores. The vendor's own examples are consulting and market research, finance, legal, life sciences and consumer research, with an illustrative prompt using a paid dataset through an app.

It fits badly where speed matters, which the vendor says plainly: "If time is critical, use search or standard chat for faster responses and save deep research for in-depth analysis." It also fits badly where the volume of research must be budgeted in advance, because no plan publishes a task allowance, and where a connector is assumed to work: "An app being available in ChatGPT or included in a plugin does not, by itself, mean it is available for deep research."