Source: raw/How_data_retention_works_when_using_Claude.md — Anthropic, “How data retention works when using Claude”, youtube.com/watch?v=PQGxYvkMobQ, 2:18, uploaded 2026-09-14. Context from the Fable 5.1 announcement, ai-research/anthropic-claude-fable-mythos-5-1-announcement-2026-09-01.md, as summarised in Claude Fable 5.1 + Mythos 5.1.
Anthropic’s short explainer of why it keeps a rolling 30 days of prompts and outputs for its most capable models, who can see that data, and what Enterprise Frontier Safeguards (EFS) changes. The core distinction is between two stores: the safety-monitoring window Anthropic keeps, and the conversation history a product keeps for you, which still follows your organization’s own settings. This is Anthropic describing its own policy; check your contract for the terms that actually apply.
Key Takeaways
- Why retention exists. Mythos-class models such as Claude Fable 5.1 bring “a major increase in intelligence and agentic capabilities”, and with it misuse ranging from fraud “all the way to sophisticated cyberattacks”. The worst misuse “can spread across many sessions and accounts”, which is hard to see one interaction at a time, so spotting it “requires storing data for a meaningful period of time.”
- What is kept. Starting with Claude Fable 5, “the last thirty days of prompts and Claude’s outputs are stored for safety monitoring.” Automated systems scan that rolling window for signals of serious misuse.
- Your chat history is separate. “Any conversation history a product keeps for you still follows your organization’s own retention settings.”
- Training. “Anthropic does not train on enterprise data without explicit permission.”
- Who can read it. By default, “no Anthropic personnel can read the data used for monitoring.” Human review happens only in rare cases, for example when an automated system flags content as potentially harmful, and only by “a small set of approved reviewers”. Any access is recorded in “a tamper-proof log.”
- Enterprise Frontier Safeguards (EFS). Described as combining “the privacy of zero data retention with state-of-the-art safeguards for detecting misuse”, developed with “more than a hundred customers”. Eligible organizations can keep the monitoring data in their own cloud infrastructure, under their own encryption keys, access policies and audit logging. When monitoring detects a pattern that needs attention, “an alert goes directly to your team with no Anthropic human review required.”
- Timing. EFS rolls out in phases “starting later this fall”. Until it is ready, eligible organizations get zero data retention on Fable 5 and Fable 5.1. This matches the 2026-09-01 Fable 5.1 announcement.
Why it matters for client work
- Two questions in a compliance review, not one: how long does the product keep my conversations (your setting), and does the provider keep a monitoring window (30 days on Fable-class models unless you are on interim ZDR or EFS).
- Agencies with client data commitments should confirm eligibility for interim zero data retention before running client material through Fable 5 or 5.1. Anthropic’s Position, August 2026 records retention as one of two explanations offered for Fable 5’s slow enterprise uptake.
- Other models. The video covers Mythos-class models only. Opus 5.5 offers zero data retention “like previous Opus models.”
Try It
- Map which models your team uses. If any workflow runs on Fable 5 or Fable 5.1, note that the 30-day monitoring window applies unless your organization is on interim ZDR.
- Ask your Anthropic account team about EFS eligibility if you are on Enterprise and need monitoring data kept in your own cloud.
- Check your organization’s own retention setting for conversation history separately; the monitoring window does not change it.
- Add the human-review terms to your vendor file: default no personnel access, rare flagged review by approved reviewers, tamper-proof access log.
Open Questions
- Who is “eligible”. The video says “eligible organizations” for both EFS and interim ZDR but does not define eligibility. The Fable 5.1 announcement names the surfaces EFS rolls out on, not the criteria.
- Consumer plans. The video describes enterprise data. Whether the 30-day window and the human-review rules apply the same way on Pro and Max is not stated.
- EFS dates. “Later this fall” and “in phases” give no dates per surface.
Related
- Claude Fable 5.1 + Mythos 5.1 — the announcement that brought back ZDR and introduced EFS.
- Mythos 5 — where the 30-day requirement started.
- Anthropic’s Position, August 2026 — the retention-versus-price debate over Fable adoption.
- Claude Opus 5.5 — the default model, with zero data retention available.
- Zero Trust for AI Agents — the wider security model for agent deployments.
- Claude for Government — audit-log and deployment controls on the public-sector side.
- Claude Output Watermarking — Anthropic’s other output-level policy change in 2026.