Anthropic shifts frontier-model retention after enterprise pushback

Anthropic revised its enterprise data retention policy for frontier models after customer pushback, replacing a mandatory 30-day retention rule with the Enterprise Frontier Safeguards program that stores monitoring data in customer-controlled infrastructure. Eligible customers receive temporary zero-data-retention access to Fable 5 and Fable 5.1, while Anthropic retains real-time classifiers and usage policy enforcement. The program rolls out in phases beginning in fall 2026 across platforms including Claude Enterprise, Amazon Bedrock, Google Cloud Agent Platform, and Microsoft Foundry.

Anthropic shifts frontier-model retention after enterprise pushback - Anthropic’s original policy required prompts and outputs from covered frontier models to be retained for 30 days for safety monitoring, including on platforms where customers had previously enabled zero data retention. <sup \ 1\ https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-models </sup - The revised Enterprise Frontier Safeguards program will store retained monitoring data in infrastructure controlled by the customer rather than Anthropic. Eligible organizations receive temporary zero-data-retention access to Fable 5 and Fable 5.1 while the program rolls out. <sup \ 2\ https://www.anthropic.com/news/enterprise-frontier-safeguards </sup - The new arrangement does not eliminate safety controls: Anthropic says real-time classifiers, its usage policy and enforcement systems still apply, and it has not published a fixed retention period for data held under EFS. <sup \ 2\ https://www.anthropic.com/news/enterprise-frontier-safeguards </sup Anthropic is changing how it handles enterprise data for its most capable AI models after customers objected to a June policy requiring 30 days of retention for safety monitoring. Under the new Enterprise Frontier Safeguards program, eligible customers can keep retained monitoring data in cloud infrastructure they control, while Anthropic continues to apply automated safeguards to model traffic. \ 1\ https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-models \ 2\ https://www.anthropic.com/news/enterprise-frontier-safeguards From mandatory retention to customer-controlled storage Anthropic’s June 9 policy covered Fable 5 and other designated frontier models, requiring prompts and outputs to be retained for at least 30 days across Anthropic’s platform and cloud partners. Data was automatically deleted after 30 days unless safety systems flagged it or the company had a legal reason to keep it. Consumer Free, Pro and Max plans were unaffected by that update because those products already retain inputs and outputs under their existing terms. \ 1\ https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-models Anthropic’s revised terms provide eligible enterprise customers with zero-data-retention access to Fable 5 and Fable 5.1 as a transition measure. EFS will roll out in phases beginning in fall 2026 and will support Claude Enterprise, Claude Code, the Claude Platform, Amazon Bedrock, Google Cloud Agent Platform and Microsoft Foundry. <sup \ 2\ https://www.anthropic.com/news/enterprise-frontier-safeguards </sup Anthropic has not disclosed a universal number of days for data retained under EFS. Safety monitoring remains part of the deal Anthropic says EFS combines automated safety monitoring with customer-controlled storage. Its current product documentation says human review is expected to be performed by the customer by default under the arrangement, rather than by Anthropic. The company also says its usage policy, real-time safety classifiers and enforcement systems continue to apply. \ 2\ https://www.anthropic.com/news/enterprise-frontier-safeguards \ 3\ https://www.anthropic.com/claude/fable The change followed direct work with more than 100 customers in financial services, healthcare, manufacturing, telecom, law, retail and the public sector. FIS, whose systems support financial institutions, said any safeguard used in its environment must meet a high standard of trust. <sup \ 2\ https://www.anthropic.com/news/enterprise-frontier-safeguards </sup Anthropic says it does not train on enterprise data without explicit permission. Companies mentioned Further sources The stories that matter, in one email. Free — unsubscribe anytime.