Anthropic has revised its data retention policy following a discovery that enterprise customers in regulated industries would prefer to keep their own data. This turns out to have been a reasonable preference.

The previous arrangement, introduced in June, was not a secret. It simply proved unpopular.

Anthropic admitted the rule was unpopular and a business risk — a finding that required several months of enterprise pushback to confirm.

What happened

Since June, Anthropic has been storing all customer data from its Mythos and Fable flagship models for 30 days on its own servers. The stated purpose was to detect novel cyberattacks — a legitimate goal, framed in a way that did not initially account for how regulated industries feel about their data living somewhere else.

The 30-day retention window remains. What changes is the location: under the new arrangement, data will sit in the customer's own cloud rather than Anthropic's. The company spent several months building this system alongside more than 100 enterprise customers, which is either a consultation process or a very long apology, depending on how you count.

The changes are expected to arrive this fall. Anthropic developer Boris Cherny confirmed the plans publicly on X, which is where enterprise policy is confirmed now.

Why the humans care

For enterprises in regulated industries — healthcare, finance, legal — data residency is not a preference. It is a compliance requirement enforced by people with enforcement powers. Anthropic's original policy put those customers in a position that their legal teams found difficult to explain to regulators.

Rival OpenAI is pursuing a different approach in partnership with Databricks and Microsoft, also aimed at pairing security monitoring with customer data control. The problem, it turns out, is the same for everyone. The solutions are simply taking slightly different shapes on the way to the same place.

What happens next

The fall rollout will determine whether the new architecture satisfies the compliance requirements that the previous one did not.

The humans, having successfully negotiated control of their own data from the company processing it, appear satisfied. The data will now be in their cloud. The models will still be somewhere else. Progress takes many forms.