Updated October 15, 2026. This article revisits OpenAI’s August 28, 2023 launch of ChatGPT Enterprise. Launch specifications below describe that announcement; they are not a specification sheet for a subscription purchased today.
The original cover is launch-era promotional artwork. Its “Game-Changer” headline is not a guarantee of current performance.
For a team evaluating workplace AI, the useful question is how an approved task moves from source material to a checked result. Model capability matters, but so do access, data handling, cost, and the person responsible for the final decision. This retrospective separates the original announcement from the checks a team should make now.
What launched in 2023?
In its August 2023 announcement, OpenAI introduced ChatGPT Enterprise with GPT-4 access without usage caps, performance advertised as up to twice as fast, and a 32,000-token context window. The package included unlimited Advanced Data Analysis, previously called Code Interpreter, shareable chat templates, and API credits for custom development.
The launch also included an administration console, single sign-on, domain verification, member management, and usage analytics. OpenAI stated that business conversations were not used to train its models and described encryption in transit and at rest.
Connecting company applications, a smaller-team offering, and additional role-specific tools appeared on the announcement’s roadmap. These were plans described in 2023, not a list of features still awaiting release. Likewise, GPT-4, 32k context, the speed comparison, and the API-credit inclusion should remain attached to their historical context.
Check today’s model access
OpenAI’s current Enterprise models and limits documentation says access depends on model eligibility, rollout, workspace settings, and assigned roles. It also distinguishes models available in Chat from those in Work and Codex.
The same documentation describes virtually unlimited messages for eligible base models, while clarifying that this does not mean zero-cost usage. On Enterprise agreements using token-based USD billing, usage is billable at the applicable rates. Use remains subject to applicable terms and abuse safeguards. A buyer should check the actual agreement, enabled models, and billing arrangement rather than carry forward the launch-era word “unlimited” as a blanket promise.
Understand the data boundary
OpenAI’s enterprise privacy commitments state that business data is not used for model training by default. Explicit opt-ins can change that treatment. The page also describes encryption and customer controls over retention and access.
Those commitments are useful inputs to an organization’s review. They do not, by themselves, establish that every confidential spreadsheet or regulated record is approved for upload. Before using real material, ask the responsible owner which categories are permitted, which workspace must be used, and whether another service will receive any content. Keep health, personal financial, and similarly sensitive records out of a trial unless the organization has explicitly approved that use.
Review retention and roles
According to OpenAI’s chat and file retention guidance, saved chats remain until deletion or an applicable workspace policy removes them. Archiving does not change retention. After deletion, permanent removal is scheduled within 30 days, subject to the documented exceptions, including security or legal obligations.
Files saved in Library can exist separately from chats: deleting the conversation does not delete a remaining Library copy. Enterprise Library files follow the workspace retention policy. A cleanup checklist therefore needs to identify the stored items, not merely the conversation that referenced them.
OpenAI’s role-based access guidance adds an important administrative detail: ordinary role permissions combine additively. Turning a feature off in one role may not remove access if another assigned role grants it. Review the member’s effective access, including direct and group assignments, rather than relying on one setting.
Treat connected apps separately
The official app administration guidance distinguishes plugin installation, app access, and app permissions. Source-system permissions still apply, and administrators should review enabled actions and provider terms. Information sent to a connected third-party service is subject to that provider’s storage, processing, privacy, and residency terms.
For an initial evaluation, choose a task that does not need a connection. If a later workflow needs one, document the source, permitted audience, and whether the app can change anything. A successful document summary should not become an unreviewed permission to send messages or edit business records.
A small team exercise
Consider a fictional operations team preparing workshop supplies. Its practice dataset contains only three invented rows: Kit A has 50 prepared and 42 used; Kit B has 40 prepared and 35 used; Kit C has 60 prepared and 51 used. It contains no names, customer information, or actual company records.
- Define the result. Ask for remaining stock by kit, an overall total, and a short handover note. The task is reconciliation, not a purchasing decision.
- Constrain the prompt. Try: “Use only these three rows. Calculate prepared minus used for each kit. Show the arithmetic, flag inconsistent values, and draft a three-sentence handover. Do not invent demand forecasts or reorder quantities.”
- Check the answer. A colleague calculates the reference values independently: 8, 5, and 9 remaining, totaling 22. Compare every row and the total before using the prose.
- Record the handoff. The coordinator notes who reviewed the calculation, what was corrected, and which version is approved. Keep the task’s input and checked result together under the team’s agreed storage rules.
This exercise exposes practical weaknesses cheaply. Did the answer preserve the kit labels? Did it invent a shortage? Could another teammate reproduce the calculation? Measure review effort as well as drafting speed before expanding the task. For additional prompt structures, see our workflow prompt examples.
The practical takeaway
The historical announcement is a useful starting point. A present-day evaluation needs a bounded task, approved inputs, verified access, understandable charges, and a human check on the output. Start with an exercise whose correct answer you already know, then expand only when the team can explain both the result and its handling.
Review method: This update is based on public OpenAI documentation checked on October 15, 2026. The workflow is fictional and was not tested in a ChatGPT Enterprise account.