Claude for Enterprise

What the Claude Enterprise plan includes in 2026 - $20 a seat plus usage, SSO and SCIM, audit logs, custom retention, and the BAA carve-outs to know.

Claude Enterprise costs $20 per seat per month plus usage billed at API rates, is sold annually only, and starts at 20 seats self-serve or 50 seats through a salesperson. What it buys over the Team plan is governance: SCIM provisioning, audit logs, the Compliance API, custom data retention, customer-managed encryption keys, US-only inference, IP allowlisting, custom roles and a HIPAA BAA option.

This page is written for the person who has to defend the purchase - the security questionnaire, the retention policy, the rollout plan. The architecture side lives on Claude enterprise solutions and deployment patterns.

#How is Claude Enterprise priced?

Enterprise is consumption-based, which is the single biggest difference from every other plan. The $20 seat fee buys platform access - admin console, identity, controls - and nothing else. All usage across Claude, Claude Code and Cowork is billed separately at standard API rates based on what people actually consume. There are no per-seat usage limits and no token allowance to run out of.

That removes the ceiling but replaces it with a forecasting problem. Anthropic publishes what it explicitly labels rough planning estimates for per-user monthly spend, which are the only official numbers available for budgeting:

Treat these as planning anchors, not quotes. The practical control is spend limits, which admins set at organization, group and user level - configure them before you invite anyone, not after the first invoice. If your usage profile is mostly light chat and your headcount is under 150, run the numbers against Team seats on the pricing page first; a fixed $20 or $25 seat with an included allocation can be cheaper than a $20 seat plus metered consumption.

Since 12 February 2026 Enterprise has been available self-serve, with a single seat type covering Claude, Claude Code and Cowork. Older Enterprise contracts may still carry separate “Chat + Claude Code” or premium seat structures - check which one you are actually on before comparing quotes.

#What does Enterprise include that Team does not?

Team is a collaboration plan with SSO and central billing. Enterprise is a governance plan. The gap is narrower than the marketing suggests in some places and much wider in others.

The row to read twice is the Compliance API. Anthropic states plainly that Team plans cannot access it. If your legal or security team needs programmatic export of activity feed events, chat data and file content across deployments - for eDiscovery, DLP or supervision - that requirement alone decides the plan. The Compliance API now extends to Cowork and, in beta, to Claude Code.

#Retention: the default is indefinite

#Identity and roles

SSO runs through WorkOS with Okta, Entra ID, Google, OneLogin, JumpCloud and Duo SAML all documented. You can verify multiple domains against one organization, but all of them must use a single identity provider. Anthropic acts as the SAML Service Provider; attribute mapping is where most configurations break. There are no passwords on Claude accounts at all - the consumer routes are Google sign-in and email magic links, which the login guide covers.

Both Team and Enterprise use Primary Owner, Owner, Admin and User roles; only Enterprise adds custom roles, where members have no default permissions and access is determined entirely by the roles assigned to their groups. Only the Primary Owner can provision new seats, request data exports and transfer primary ownership - worth knowing before that person goes on leave.

#What are the HIPAA and BAA carve-outs?

Anthropic offers a Business Associate Agreement covering its HIPAA-ready services, including Enterprise plans and the first-party Claude API, activated by the Primary Owner under Organization settings. The coverage is genuinely useful and genuinely incomplete, and the exclusions are not obvious from the product surface.

Two further operational traps. Organizations with zero data retention cannot use Claude Code cloud session features at all. Organizations with IP allowlisting will see Anthropic-hosted cloud sessions fail authentication unless those ranges are exempted. Both are fixable, but only if you find them before the pilot rather than during it.

#Defaults that differ on Enterprise

Anthropic ships several surfaces in a more conservative state for Enterprise organizations: Cowork cloud sessions default off (they default on for Team), Claude Design defaults off, Claude in Chrome requires an admin to enable it, and Claude Science needs admin approval. Since 1 July 2026 admins have also controlled which models and which thinking effort levels their organization can use. The full per-plan picture is in the feature and plan matrix.

#Does Anthropic train models on enterprise data?

No. The Commercial Terms, which govern Team, Enterprise and API use, state that “Anthropic may not train models on Customer Content from Services” and that the customer “retains all rights to its Inputs” and “owns its Outputs”. Anthropic's privacy centre restates it: by default, inputs and outputs from commercial products are not used to train models.

There is one carve-out worth telling your users about. If someone explicitly submits feedback - the thumbs up or thumbs down button - that conversation may be used for training and is retained for up to five years. Team and Enterprise owners can disable this entirely through the “Rate chats” setting under Data and Privacy. The consumer training rules are completely different and do not apply here; both regimes are set out with clause language on commercial use rights.

#What does a realistic Claude Enterprise rollout look like?

Six weeks is achievable for an organization that already has an identity provider and a decision-maker. The failure mode is not technical - it is discovering a compliance requirement in week five that needed a different configuration in week one.

The organizations that stall are the ones that treat week 3 as a launch. A pilot cohort exists to find the two connectors nobody anticipated and the one team whose data cannot leave a jurisdiction. Deployment architectures for the phase after week 6 - Bedrock, Vertex AI, Foundry, agent patterns, governance - are on enterprise solutions, and the automation patterns behind them on automating work with Claude.

#What should procurement ask before signing?

A checklist that reflects how Claude Enterprise is actually configured, rather than a generic SaaS questionnaire:

• Retention. What period will we set, and who owns changing it? Remember the default is indefinite and the floor is 30 days.
• Compliance API. Do we need it? If yes, Team is not an option, and Public Sector organizations are excluded.
• BAA scope. Which surfaces will handle PHI? Confirm Cowork is out, and that Claude Code needs ZDR in a supported mode.
• Excel. Do we allow the add-in knowing it sits outside custom retention, audit logs and the Compliance API?
• Zero data retention. If we need it, are we accepting the loss of Claude Code cloud sessions?
• IP allowlisting. If we enable it, have we exempted Anthropic-hosted cloud sessions - or accepted that they will fail?
• Inference region. Do we require US-only inference, and does that conflict with any regional obligation elsewhere?
• Identity. One IdP for all verified domains - does that work for subsidiaries and acquisitions?
• Spend. What are the org, group and per-user limits on day one, and who gets alerted?
• Feedback. Do we disable “Rate chats” to close the five-year feedback retention path?
• Model access. Which models and effort levels are permitted? Restricting to Opus 5 and Sonnet 5 is a reasonable starting policy.
• Exit. Who is the Primary Owner, who is the backup, and how do we request a data export?

One more item that is easy to miss: Claude Security, in public beta for Enterprise, scans GitHub-hosted repositories for vulnerabilities and proposes patches. Anthropic charges direct token cost only, with no additional platform fee - but it may require allowlisting the range 160.79.104.0/21, which is a firewall change someone has to own.

#Frequently asked questions