If your enterprise signed a zero-data-retention (ZDR) agreement with Anthropic and assumed it covered every Claude model you'd ever deploy, the covered-models policy that took effect June 9, 2026 changes that assumption for the two models you're most likely to want: Claude Fable 5 and Claude Mythos 5. Anthropic now stores every prompt and every output from those models for 30 days — no exceptions, no opt-outs — and the August 2026 news cycle sharpened the contrast when OpenAI previewed Private Safety Processing on the opposite path: keep ZDR, send only a narrow safety signal.
This post is the Anthropic-side explainer enterprise security and compliance teams actually need — not a rehash of OpenAI's announcement, but what changed in your contract posture, what CMEK buys you, and what you can do today if regulated data cannot sit on a third party's servers for a month.
TL;DR: what people are asking
| Question | Answer |
|---|---|
| Which models require 30-day retention? | Claude Fable 5 and Claude Mythos 5 (Anthropic's "covered" / Mythos-class frontier models). |
| Does ZDR still exist for these models? | No — prior ZDR agreements are overridden for covered models on all platforms. |
| Is retained data used for training? | Anthropic says no — safety monitoring only. |
| Can I opt out? | No opt-out for covered models. Use older Claude versions with ZDR for sensitive workloads. |
| Does CMEK exempt me from retention? | No — CMEK controls encryption keys, not the retention window. |
| How does this differ from OpenAI? | OpenAI keeps ZDR and sends automated signals; Anthropic retains content for human review of flagged sessions. |
| Where is the official policy? | Anthropic Help Center — Data retention for Covered Models |
What actually changed on June 9, 2026
Anthropic's covered-models data retention policy applies universally:
- Every platform: Claude.ai consumer, Teams, Enterprise, Claude API, AWS Bedrock, Google Cloud Agent Platform, Microsoft Foundry.
- Every interaction: prompts submitted to and outputs generated by covered models.
- 30-day window: automatic deletion after 30 days unless flagged for safety investigation or legal hold.
The policy exists because single-session safety checks miss threats that unfold across many requests — multi-turn jailbreaks, coordinated abuse across accounts, or an agent that keeps acting after being told to stop. Anthropic's August 2026 risk report acknowledged the business risk explicitly: "We have recently announced our plan to require 30-day data retention on our most capable models — a decision we believe will be unpopular with customers who have come to expect zero retention."
For teams that built compliance programs around ZDR on Claude Fable 5 and Mythos 5, that sentence is the headline.
What enterprise controls you still have
Anthropic is not leaving enterprises with zero knobs — but the knobs control how data is protected, not whether it is retained.
Customer-managed encryption keys (CMEK)
Eligible Enterprise and Claude Platform organizations can enable CMEK: you provision a key in AWS KMS, Google Cloud KMS, or Azure Key Vault; Anthropic encrypts workspace data at rest with it. You retain rotation, audit, and revocation control; key operations appear in your cloud provider's audit logs.
What CMEK does not do: shorten or eliminate the 30-day retention requirement for covered models.
Access transparency and tamper-proof logs
Anthropic's documentation states that by default no Anthropic personnel can read retained conversations. Human review happens only when automated trust-and-safety systems flag content, through a controlled access path limited to approved reviewers. Every access instance is recorded in a tamper-proof log reviewers cannot suppress.
For legal and compliance teams, that audit trail matters — but it does not satisfy regimes where any third-party storage of plaintext prompts is unacceptable.
Sandbox organizations
For Enterprise customers with ZDR today, Anthropic offers setting up a separate sandbox org to test covered models under the new retention rules before touching production — configured through the admin console or via Anthropic support.
What this means for what you build or pay
| Workload type | Practical recommendation |
|---|---|
| Regulated data (HIPAA, defense, M&A diligence) | Route to Claude versions that still support full ZDR, or air-gap sensitive prompts before they reach Fable 5. |
| General coding / product work on Fable 5 | Accept 30-day retention; enable CMEK if your org qualifies; document the safety-only, no-training commitment in your vendor risk file. |
| Multi-tenant SaaS built on Claude API | Update your privacy policy and DPA — downstream customers may assume ZDR applies to all models. |
| Model selection / procurement | Treat retention policy as a first-class eval criterion alongside benchmark scores — same lens as choosing open-weight vs closed models. |
OpenAI vs Anthropic: two safety architectures
The August 2026 news cycle paired both announcements:
| Dimension | Anthropic (covered models) | OpenAI (Private Safety Processing preview) |
|---|---|---|
| Default for frontier models | 30-day content retention | Zero retention maintained |
| Cross-session abuse detection | Human review of flagged retained sessions | Automated pattern detection on customer-controlled or customer-key-encrypted content |
| What OpenAI/Anthropic sees | Flagged session content (limited reviewers) | Narrow categorized signal only (OpenAI path) |
| Enterprise friction | High for ZDR-dependent customers | Lower for existing ZDR customers |
| Availability | In effect since June 9, 2026 | Preview; broader rollout + white paper planned September 2026 |
Neither company eliminated safety monitoring. They made different bets on whether catching sophisticated multi-turn abuse requires retaining plaintext or can be done with encrypted processing and signals.
What people are asking in security reviews
"Is this a GDPR problem?" Retention with a defined purpose (safety) and 30-day deletion is easier to document than indefinite storage — but you still need a lawful basis and must update subprocessors and DPAs if customer data now sits on Anthropic servers longer than your prior ZDR contract assumed.
"Can we use Bedrock/Foundry to avoid this?" No. The policy applies on third-party platforms too.
"Should we downgrade to Sonnet for sensitive work?" If an older model still supports ZDR and meets your quality bar, that's the straightforward split: frontier models for non-sensitive velocity, ZDR-eligible models for regulated inputs. See Claude Code pricing for model-tier cost tradeoffs.
"Did Anthropic shift log custody to enterprises?" Partially misreported. Anthropic did not hand log custody to customers — it requires retention and offers CMEK plus access-transparency tooling so enterprises control encryption and audit, not the retention decision itself.
Related on explainx.ai
- OpenAI Private Safety Processing — the opposite trade: ZDR plus automated signals
- Claude Fable 5 and Mythos 5 launch — what covered models actually are
- Choosing open-weight vs closed AI models — when retention policy pushes you toward self-hosted options
- Claude Code pricing guide — model tiers and cost
- Agent skills security threats — another enterprise trust surface
- What is MCP? — protocol context for tool-connected agents under new retention rules
Official sources: Anthropic — Data retention for Covered Models · CMEK documentation · Anthropic August 2026 risk report
Policy details, model lists, and CMEK eligibility are accurate as of August 20, 2026. Verify current terms in Anthropic's Trust Center before updating production compliance documentation.