Claude Team and Enterprise Operations Guide

Analysis

The three September decisions

Most Claude updates look like features to users and defaults to admins. This issue has three that need an explicit decision rather than a wait-and-see. First, memory: on managed plans it is admin-gated and off until a user enables it, so the answer to "does Claude remember our conversations?" depends on a switch you control. Second, Claude in Chrome: generally available on every paid plan, with domain allowlists as the Enterprise control. Third, Fable 5.1 data handling: EFS customer-controlled storage is phasing in through the fall, with zero data retention as the interim path for eligible customers.

None of these generate a support ticket until after a user has already relied on the default. Decide now, write it down, and tell users what you decided.

• Record the current memory setting for each workspace and the rationale.
• Configure or deliberately decline Chrome domain allowlists.
• Confirm EFS or ZDR eligibility with Anthropic before routing regulated data to Fable 5.1.
• Name an owner for model entitlement changes so a new release does not surprise a controlled workspace.

Analysis

Model entitlements after a four-model board

The generally available lineup is now Fable 5.1, Opus 5, Sonnet 5, and Haiku 4.5, with Fable 5 and the Opus 4.x line as legacy. Entitlement policy written against the July board will name models that are no longer current and omit the two that most users will ask for.

There is also a behavioral wrinkle worth documenting for support: Opus 5 introduced automatic fallbacks when safety classifiers flag Opus 5 or Fable 5, and biology-related Fable blocks route to Opus 5. A user who sees a different model answer than the one they selected is not necessarily seeing a bug.

Analysis

Compliance and provenance

The Compliance API is not a generic export promise for every Claude surface; the support article documents availability conditions and exclusions. Verify the exact org context before designing retention, discovery, or audit workflows around it. The failure mode is not missing data — it is an internal process built on a feature this tenant cannot use.

Text watermarking adds a provenance dimension to the same conversation. Future Claude models watermark generated text, with a detection API described as coming and a transition period for models published before August 2, 2026. Policies should treat the watermark as evidence that text is model-generated, not as attribution to a specific user, because Anthropic states it is not traceable to a person, organization, or conversation.

Analysis

Programs with their own eligibility rules

Two access programs opened since the last issue and both will reach admins as requests rather than as rollouts. Claude for Teachers gives verified US K-12 educators free premium access and is FERPA-configured, with identifiable student records requiring school or district authorization; a school and district offering launched August 28, 2026. The scientists program opened 10,000 seats, free at the standard tier and $15 per month for premium with 5x usage, for one year.

If your organization touches education or research, expect individual sign-ups before institutional agreements. That order is the governance problem, not the programs themselves.

Analysis

Incident playbook

Claude Status belongs in the admin workflow. When users report missing models, changed refusals, or unexpected errors, check status, app release notes, and platform release notes before starting a bug hunt. Anthropic’s July 30 disclosure of four cybersecurity evaluation incidents, with a METR review, is also a useful precedent for how the company communicates safety events — public post first, detail in the assessment.