ChatGPT Sites Enterprise Update: Regions and Tunnels

The ChatGPT Sites Enterprise update introduces an admin preview, regional hosting for new Sites in US/EU data-residency workspaces and tunnels to private services. OpenAI’s October 15 release notes describe a rollout to eligible workspaces, with public publishing still off by default. The tunnel rollout also targets eligible Enterprise workspaces, rather than every account. The change gives administrators three separate decisions: member access, hosting location and access to an internal service.

The operational instructions below explain those decisions using the linked current documentation. They are not a claim that every existing Sites control first launched on October 15. Our DevDay availability report covers the earlier broader product announcements.

ChatGPT Sites Enterprise update: inspect preview and launch access

The workspace administration guide distinguishes current preview access from launch access. Where Sites appears in Admin Preview, administrators use Set launch access or Edit launch access to choose the launch setting. To keep it off, set the relevant Launch access choices to Off, save and reopen the setting. Follow the workspace’s launch notice for its deadline; the guide does not supply one universal date.

Public publishing requires both the workspace setting and the appropriate role-based creation and publishing permissions. An administrator checking a member’s problem should confirm that Sites is enabled, the member has an eligible role or group, and the member is signed into the intended workspace.

Connected-app use is another setting: an available ChatGPT plugin can remain unavailable in Sites. Enterprise plugins default off for Sites when no administrative choice has been saved. Tenant approval, where applicable, and each visitor’s own authorization remain separate. For connected apps, the Site must be workspace-private and the visitor must belong to that workspace. The earlier plugin administration report explains the general console workflow; this update concerns the Site consuming those capabilities.

Identify storage, hosting and inference separately

OpenAI’s residency documentation gives Sites a specific scope: new Sites inherit supported US or EU workspace storage and hosting regions. Existing Sites are not automatically migrated. Other workspace residency regions do not currently extend that location setting to Sites storage and hosting.

Workspace owners and administrators can inspect the configured region on ChatGPT web: switch to the intended workspace, open Workspace settings, select General and find Data Residency Region beside the organization and workspace identifiers. Inference Residency Region, when displayed, is separate. These fields show configuration; they are not migration controls.

Inference residency concerns in-scope GPU execution, not every processing step. Authentication, routing and other CPU processing can still occur elsewhere. External integrations follow their providers’ own storage and processing terms.

For a concrete review record, we recommend listing the workspace’s displayed region and the particular Site’s creation date together. Keep the private service’s location in a separate row. This proposed record prevents a team from substituting one regional label for three different questions. It does not establish compliance or move an existing deployment.

A private tunnel does not make the Site private

The tunnel guide describes a client run by IT on a computer or server that can reach the approved service over HTTP or HTTPS. The client carries requests and responses; the service’s authentication and access controls still apply. A tunnel does not open the entire network.

Allow members to attach private tunnels must be enabled for both the Site owner and the person connecting the tunnel. Public sharing needs an additional permission for both the owner and publisher, as well as workspace public-publishing enablement. An already-public Site must be made private before attaching a new tunnel.

Connection and operation failures require different checks. If the client is not ready, IT checks its process, credentials and reachability. If a service action fails despite a connected tunnel, inspect that service’s authorization and preserve the response code or request identifier.

Removing one Site’s connection and stopping a shared tunnel client have different reach. Stopping the client can affect other Sites using it. Revoking a runtime key may not stop an already-running connection. Keep credentials out of browser-side content and support records.

Review the published application, not just its preview

The member guide distinguishes saving a version from publishing: every deployment URL is a production URL. Save a version when reviewing changes without updating the live Site. Sending edits from a published Site starts an editing workflow; it does not immediately publish those edits.

Publication can change the database before the new version becomes live. If publishing fails, database changes may already exist. Restoring an earlier version restores code, not the database or uploaded files. Regional-hosted Sites also have a particular secret-handling distinction: changes to existing secret values can affect the live Site without another publish, while adding or removing configuration keys still requires publication.

Editors can inspect the live database and publish code using the Site’s secrets and storage. After the owner’s first publication, an editor can publish later versions without a separate owner-approval step. Give that role to trusted collaborators.

Enterprise running and hosting charges are currently waived; building and editing still count toward Codex usage, and usage and storage limits remain. A hosting waiver is not an unlimited development allowance.

For model availability and subscription choices, use our GPT-6 access guide. Readers comparing access products should keep that overview alongside this dated, administrator-focused release report.