Anthropic just drew a firm line. Starting June 9, 2026, prompts and outputs sent to its latest high-capability Claude models face mandatory 30-day retention. No exceptions. No zero-data-retention option. The policy applies across consumer apps, enterprise workspaces, and every third-party platform from Amazon Bedrock to Microsoft Foundry.
The move targets what the company calls Mythos-class models. These include the newly released Claude Fable 5, now generally available, and Claude Mythos 5, offered under limited access to approved partners. Both fall under a new “covered models” category for systems with advanced abilities that demand extra safety scrutiny. Claude Help Center spells it out plainly: prompts submitted to, and outputs generated by, Mythos-class models are retained for 30 days for trust and safety purposes, on every platform where these models are offered.
Why now? The models pack enough power to tackle complex, long-horizon tasks. They also introduce fresh risks. Anthropic wants human reviewers to examine interactions for novel jailbreaks, multi-request attacks, and other sophisticated threats that automated systems might miss. The data helps reduce false positives too. But it comes at a cost to privacy-sensitive customers who had grown accustomed to stricter deletion promises.
Enterprise users feel the shift most. Previous zero-data-retention agreements, a selling point for handling sensitive legal, medical, or financial information, do not apply here. Requests to these models from organizations without compliant retention settings return errors. One LinkedIn post from AI consultant Melinda Byerley captured the reaction in enterprise circles: no zero data retention policy available in your Enterprise plan for Fable at this time. Pro, Codex, and Max remain unaffected.
Anthropic insists the retention serves a narrow purpose. The company won’t use this data to train new models. Access receives strict controls. Every human review gets logged. After 30 days the information deletes automatically unless tied to an active safety probe or legal hold. Those protections appear in the official support article, yet some customers still see red flags.
Reddit threads lit up within hours of the announcement. One popular post titled “Mythos / Fable 5 REQUIRES data retention!” warned users to watch what they share. Comments ranged from acceptance of the safety trade-off to outright frustration over broken expectations for confidential work. The discussion highlighted a growing tension. As models grow more capable, labs face pressure to monitor them closely. Users, especially in regulated industries, want ironclad guarantees their data vanishes quickly.
This isn’t Anthropic’s first adjustment to retention rules. Last year the company updated consumer terms so users who opt in for model improvement see chats kept up to five years in de-identified form. Those who opt out stick with the shorter window. The Mythos policy differs because retention is compulsory and limited strictly to safety, not training. Still, it signals a broader trend. Frontier labs increasingly view some level of persistent data access as essential for responsible deployment of their most powerful systems.
Claude Fable 5 brings notable performance gains. Early reports position it as a step above previous Opus-class models for agentic coding, deep document analysis, and high-stakes reasoning. Pricing sits at $10 per million input tokens and $50 per million output tokens. Availability spans Claude’s own interfaces plus major cloud marketplaces. Yet the retention requirement has already prompted some developers to weigh alternatives or limit usage to non-sensitive tasks.
Third-party platforms must adapt too. GitHub noted that unlike other Claude models in Copilot, Claude Fable 5 requires data retention to operate Anthropic’s safety classifiers. Microsoft, Google, and Amazon face similar updates for their customers. A recent analysis on Jessica Eaves Mathews’ Substack laid out the implications clearly. Existing zero-data-retention deals do not cover Fable 5 traffic. The policy overrides prior commitments for this model class.
Privacy advocates question whether 30 days strikes the right balance. Short enough to limit exposure, perhaps. Long enough for thorough review of complex attack chains. But what counts as a safety investigation? How many employees gain access? The support page promises logging and deletion in almost all cases. Details remain thin. Anthropic has not published a full transparency report on reviewer numbers or false-positive rates for these new classifiers.
Compare this to earlier Claude releases. Standard models offered 30-day default deletion with opt-in training or zero-retention paths for qualifying enterprise buyers. Covered models close those doors. The Covered Models page lists Claude Mythos 5 with a June 9, 2026 release and limited availability. Fable 5 appears as generally available across more surfaces. Both carry the 30-day minimum retention by default. Zero data retention is unavailable.
Industry watchers see this as part of a larger pattern. As AI capabilities cross certain thresholds, providers accept that full anonymity becomes harder to guarantee. A 2026 system card for an earlier Mythos preview discussed advanced cyber capabilities that emerged from general improvements in code reasoning rather than targeted training. Those findings led to restricted initial access through Project Glasswing. Now the company releases broader versions while tightening data oversight.
But. The decision risks alienating exactly the customers most likely to push these models into production. Law firms, hospitals, banks. Organizations that once praised Anthropic’s privacy stance may hesitate. Some will route sensitive work to older models that still support zero retention. Others may explore competitors. A post on Coursiv noted that Fable 5 does not support zero data retention, a point repeated across developer forums.
Anthropic argues the policy enables safer innovation. Better monitoring means faster detection of novel risks. Fewer false positives mean less disruption for legitimate users. The company has invested heavily in constitutional classifiers and other safeguards. Retention supplies the human-in-the-loop element for edge cases that automation cannot yet handle reliably.
Even so, the optics matter. Enterprises negotiating AI contracts now add new clauses about model-specific retention. Legal teams scrutinize which versions fall under covered status. Procurement departments ask hard questions about data flows before approving spend. The era of simple plug-and-play frontier models appears to be ending.
Recent X conversations reflect the split. Developers celebrate Fable 5’s coding prowess while flagging the retention warning. Security researchers debate whether the 30-day window actually improves safety or simply creates a larger target for breaches. One thread noted that no anonymous usage exists for these covered models anymore. Prompts tie back to accounts on every platform.
Longer term, this policy could influence regulation. Lawmakers watching AI safety may view mandatory retention for high-risk models as a reasonable step. Privacy regulators could push back, demanding stricter deletion timelines or independent audits. California’s existing rules on frontier models already require certain disclosures. Future bills might codify data-handling standards for systems above specific capability levels.
Anthropic has promised to evolve the approach. Future updates may shorten retention for lower-risk interactions or improve automated detection enough to reduce human review needs. For now the rule stands clear. If you use Claude Fable 5 or Mythos 5, your inputs and outputs stay in Anthropic’s systems for 30 days. Plan accordingly.
The company faces a classic frontier-AI dilemma. Build systems powerful enough to transform industries. Then accept that those same systems require closer watching than anything that came before. Retention isn’t perfect. It isn’t anonymous. But in Anthropic’s view it represents the minimum needed to deploy these models responsibly. Customers will decide whether that calculation holds.
WebProNews is an iEntry Publication