ChatGPT Enterprise vs Claude for Work: the data and admin comparison

Most comparisons line up ChatGPT Enterprise against Claude Enterprise as though both sat on a shelf with a price tag. They do not. One is listed publicly at $20 a seat with a twenty-seat floor. The other has no published price, no published minimum, and arrives as a negotiated invoice. For a company of 50 to 500 people, that asymmetry settles more than any capability difference between the two assistants.

The second thing those matrices hide is that neither vendor sells one product. Claude for Work is a Team plan and an Enterprise plan. ChatGPT for business is also two, Business and Enterprise. Nearly every control a security reviewer asks about sits on the upper tier of both. So the useful question is not which assistant is better. It is which tier of which vendor clears your own list, at a price you can see.

The tier line is the decision

Both vendors structure their business plans the same way. The lower tier gets a shared workspace, central billing, single sign-on and a usage dashboard. The upper tier gets the governance surface: directory-driven provisioning, role-based access, an exportable log, and a retention window you set rather than inherit.

What differs is how you reach the upper tier. Anthropic publishes Claude Enterprise at US$20 per seat per month, billed annually, with a twenty-seat minimum, and notes that usage cost scales with the model. A 60-person company can read that number, model it, and buy it. OpenAI publishes nothing for ChatGPT Enterprise: the plan is quote-only and prepaid annually. The figures circulating in procurement write-ups, roughly $45 to $75 a seat against a 150-seat floor, come from third parties rather than OpenAI, and we would not put them in a board paper without a quote.

That reorders the comparison for a mid-market buyer. If your list includes automatic de-provisioning and an exportable audit log, one vendor sells that at a price you can look up today and the other sells you a conversation. Comparing the top tiers as equals implies a symmetry that is not there.

What a seat costs

Every figure below was read off a vendor page or help center in October 2026. Both catalogs move, so check before quoting it internally.

	Claude Team	ChatGPT Business	Claude Enterprise	ChatGPT Enterprise
Standard seat, annual	$20/mo	$20/mo	$20/mo + usage	Not published
Standard seat, monthly	$25/mo	$25/mo	Annual only	Annual only
Premium seat, annual	$100/mo	$100/mo	n/a	n/a
Seat minimum	2	2	20	Not published
Single sign-on	Yes	ChatGPT only	Yes	Yes
SCIM provisioning	No	No	Yes	Yes
Audit log	No	No	Yes	Via Compliance API
Retention control	No	No	Yes, 30-day floor	Yes, 90-day floor
Compliance API	No	No	Yes	Yes

On published prices a standard seat costs the same on both sides, $20 a month annually or $25 monthly, with premium seats at $100 and $125. Headline seat price has stopped being a differentiator. Anyone selling you a decision on it is comparing two identical numbers.

The commercial shape does differ at the top. Claude Enterprise bills a cheap seat plus usage at API rates, pooled across the organization, with spend limits an admin sets per person and company-wide. ChatGPT sells a flat seat. Flat seats forecast more easily; metered usage costs less when half your staff open the tool twice a week. Which is cheaper depends on a usage spread you will not have until month three, an argument for a short pilot over a long negotiation.

Provisioning: the question a reviewer asks first

Here is the asymmetry that matters more than price and appears in almost no comparison: ChatGPT Business has no SCIM and no directory group sync. OpenAI’s help center says so plainly, and adds that Business single sign-on reaches ChatGPT only, not the developer platform. Provisioning and de-provisioning on that plan are manual.

Translate that into an offboarding run. Someone leaves on a Friday. Your identity provider disables the account and every service wired to it follows. The ChatGPT Business workspace does not, because nothing told it to. An admin removes the seat by hand, from a list, on a day when four other things are also happening. Claude Team has the same gap: single sign-on and domain verification on Team, SCIM only at Enterprise.

At the lower tier, then, both vendors leave you running a manual process against human memory. What differs is the cost of closing it: a listed plan at twenty seats on the Claude side, a negotiated agreement on the ChatGPT side. If you have promised anyone evidence that access was revoked, this row belongs at the top of your evaluation. It is the first thing a customer security questionnaire asks.

Logging: what you can hand an auditor

Both top tiers log, in different shapes, with different retention.

OpenAI’s Compliance API, now part of a wider Compliance Logs Platform, exports time-stamped records of conversations, uploaded files, workspace GPT configuration, memories and workspace users, with admin audit and authentication logs added recently. Exports arrive as immutable, time-windowed JSONL, and eDiscovery and data-loss-prevention vendors including Microsoft Purview have built integrations. One number matters for planning: the conversation log is a rolling 30-day window. A retention duty longer than a month means you own the archive, not OpenAI.

Claude Enterprise lists audit logs, a Compliance API, usage reporting and OpenTelemetry monitoring. The OpenTelemetry piece is the quietly useful one if you already run a security event platform: telemetry lands in a format your pipeline reads instead of an integration you maintain.

Both satisfy a reasonable auditor. Neither is enough alone: a log covers the sanctioned tool and says nothing about the personal account someone used instead.

Retention: four numbers nobody prints

We read ten comparison pages for this article. Not one printed a retention window. All four numbers below sit on vendor help pages.

• ChatGPT Business: indefinite, not configurable. Chats, files and canvas documents are kept until a user deletes them. There is no workspace setting.
• ChatGPT Enterprise: a policy you choose. Owners set a workspace retention policy, 90 or 180 days and upward, and member removal follows it. Files not saved to the Library expire after 48 hours.
• Claude Enterprise: indefinite by default, 30-day floor. Owners set a custom period in organization settings, the shortest allowed window is 30 days, and chats inside a project follow the project’s schedule rather than the standalone one.
• Deleted conversations on both: scheduled for removal within 30 days, unless law or an abuse investigation requires longer.

Two caveats for a contract review. Anthropic notes that content flagged by its automated trust-and-safety systems is held outside your custom policy, and that custom retention does not reach every surface in the product family. OpenAI’s equivalent is the rolling log window above. Neither is a scandal, and both turn a clean policy statement into a messy one if you learn it during an incident rather than during procurement.

The plain reading: buy a lower tier on either side and your retention answer is “indefinite, by default, with no setting.” Write that down for whoever asked about data handling. They will assume otherwise.

Where the two are the same

Shorter than vendor marketing suggests, and worth knowing.

Neither trains on your business data. Both exclude organizational inputs and outputs from model training by default. On the consumer products training is a user-controlled setting, so the plan you bought answers the objection, not a toggle an employee has to remember.

Both carry the baseline certifications. SOC 2 Type 2 on both sides, with Claude Enterprise’s page also listing ISO 27001, GDPR and CCPA alongside a HIPAA-ready offering. Neither is the secure one on a badge list.

Both gate connectors behind admin approval. The risk travels with the connector, not the chat window, so review the integration you switch on.

Both expose a usage dashboard that will mislead you. Seats assigned, messages sent, weekly actives. Those numbers rise on their own for a month and say nothing about whether anyone does better work, the case we make in our argument that buying licenses is not the same as building capability.

What we would buy at 50 to 500 people

For a company with no dedicated AI owner and a list of controls someone has asked about:

1. Write the data rule before you shop. One page: what may never be pasted anywhere, what may go into the company account, and what to do within an hour of a mistake. The rule decides which tier you need. Shopping first lets a feature list write your policy.
2. Decide whether automatic de-provisioning is optional. If it is not, you are buying an upper tier, and the published-price path is Claude Enterprise at twenty seats.
3. Pilot two teams on one vendor for six weeks. Not a bake-off across both. One configured workspace, on real work. The usage spread decides whether flat seats or metered usage is cheaper.
4. Configure before you announce. Connector approvals, retention window, roles, log export, in that order, while the workspace is still empty. The case for a company seat over a personal login is in our comparison of paid accounts against consumer logins.
5. Then train, on your own documents. The controls constrain the tool. A half day of hands-on work on the team’s real files moves the behavior no control reaches.

What we would not do: buy the top tier because a vendor named a seat minimum, evaluate both for six months, or treat a signed order form as the end of the governance task. Teams inside a regulated review need a different sequence, covered in the compliance training options for regulated teams.

The part no admin console covers

LayerX’s 2025 enterprise data-security research found that 77% of employees paste data into generative AI prompts, and 82% of those pastes come from unmanaged accounts outside any oversight, averaging 14 pastes a day. Those are vendor telemetry figures, not a government survey, so read them as directional. The direction is not in dispute.

Set that against adoption. The US Census Bureau’s Business Trends and Outlook Survey, collected between December 2025 and May 2026, found 37% of firms with 250 or more employees and 32% of firms with 100 to 249 employees reported using AI. Roughly a third of companies your size have the tools, and nothing in either console reaches the employee who opens a personal account on a phone because the sanctioned one felt slow.

That is a training problem wearing a procurement costume, the gap described in the playbook on policy, training and the distance between them. The tier buys revocation, evidence and a retention window. Getting staff to prefer the sanctioned account requires that it be the one they are fluent in, the job of AI security and risk awareness training for employees. It is also why whoever holds tool approvals needs their own grounding, compared in the governance training options for leaders.

How to tell at ninety days whether it worked

Set these tests before the rollout. Each is harder to run honestly later.

• Offboard a test account through your identity provider, then check the workspace an hour later. If the seat is still live, you bought a manual process.
• Ask for 60 days of conversation logs. A window shorter than your duty means you need an archive job, not another plan.
• Ask three employees separately to state the data rule. Three different answers means the policy exists on paper only.
• Count how many people used the company account for their hardest task last week. Not logins. The hardest task. That number tells you the sanctioned tool won.

If the logs come back clean and the third test fails, the purchase worked and the AI enablement did not. That is the common outcome, and cheaper to catch now than at renewal.

Frequently asked questions

Is Claude for Work the same as Claude Enterprise?

No. Claude for Work covers both business plans. Team starts at two seats with single sign-on, domain verification, central billing and connector controls. Enterprise adds role-based access, SCIM, audit logs, a Compliance API and custom retention controls. Comparison pages that name only the top tier are how buyers end up weighing a $20 Team seat against a quoted ChatGPT Enterprise seat.

How much does ChatGPT Enterprise cost per seat?

OpenAI does not publish it. The plan is quote-based, negotiated per organization, and billed as a prepaid annual invoice. Third-party procurement write-ups report $45 to $75 a seat and a 150-seat floor; those numbers are not from OpenAI and should not reach a budget without a quote. ChatGPT Business is published at $20 a seat per month annually, $25 monthly.

Do either of them train on our business data?

Not on either vendor’s business plans. Both exclude organizational inputs and outputs from model training by default. The distinction to watch is the consumer product, where training is a user-controlled setting. That is why a personal login carrying company work is a different risk from a company seat.

How long is our data kept on each plan?

ChatGPT Business keeps chats, files and canvas documents indefinitely with no workspace setting. ChatGPT Enterprise owners set a policy from 90 days upward, and Library-excluded files expire after 48 hours. Claude Enterprise keeps data indefinitely by default with a configurable floor of 30 days. Deleted conversations on both are scheduled for permanent removal within 30 days, barring legal or abuse holds.

Which plan can IT de-provision automatically?

Only the upper tiers. ChatGPT Business has no SCIM or directory group sync, so provisioning and de-provisioning are manual, and its single sign-on does not extend to the developer platform. Claude Team is the same, with SCIM arriving at Enterprise. If automatic revocation is a control you promised someone, you are buying an upper tier.

Do we need the top tier at 150 people?

Not automatically. Three questions decide it: must you prove access was revoked, must you produce conversation records, must you set your own retention window. If none applies, a lower tier for a year is defensible and reversible. If one does, the lower tier leaves you a manual process and an indefinite retention answer.

Does buying the governance tier satisfy an AI use policy?

No. The tier gives you the enforcement surface: revocation, logging and retention. A policy also needs a written data rule, a named owner with hours in their week, and staff who can state the rule without looking it up. The measured failure mode is pasting into unmanaged accounts, and no admin console reaches that.

What should we configure before we train anyone?

Connector approvals, the retention window, roles and the log export, in that order, while the workspace is still empty. Then run the first session on the team’s own documents rather than a vendor demo, so the sanctioned account is the one people are fluent in.

Key takeaways

• Standard seats cost the same on both sides: $20 a month on annual billing, $25 monthly, premium seats at $100 and $125. Seat price no longer separates them.
• The decision sits at the tier line. Claude Enterprise is published at $20 a seat with a twenty-seat minimum; ChatGPT Enterprise has no published price or minimum and arrives as a negotiated invoice.
• Neither lower tier has SCIM, so leavers are removed by hand on both Claude Team and ChatGPT Business.
• Retention: indefinite and unconfigurable on ChatGPT Business, a 90-day floor on ChatGPT Enterprise, a 30-day floor on Claude Enterprise. OpenAI’s compliance logs run a rolling 30-day window, so a longer duty means you own the archive.
• The exposure no control reaches is behavior: 77% of employees paste into generative AI prompts and 82% of those pastes come from unmanaged accounts.

If you are weighing a published Claude Enterprise seat against a quoted ChatGPT Enterprise agreement, or working out whether your current plan can revoke access and produce a log, book a free AI readiness call. We will go through the controls you have promised someone, which tier delivers each, and what your team needs to learn before the seats go out.