A Mandatory Rule With No Exceptions
In June 2026, Anthropic told every enterprise customer running its most capable models that every prompt and every response would be logged for 30 days, with no exceptions and no opt-out, regardless of what any earlier contract promised.
Anthropic framed the rule as a safety measure, a way to detect misuse and cyberattacks carried out through its Claude Fable 5 and Claude Mythos 5 models. For a bank overseen by the UK's Information Commissioner's Office, the practical effect was the same either way: prompts that might contain client records, unpublished source code or trade secrets would now sit on Anthropic's own servers for a month, with no way to say no.
What Actually Changes
On 20 August 2026, Anthropic told Bloomberg and Reuters it will let enterprise customers keep those mandatory 30-day logs on infrastructure they control instead of Anthropic's own cloud, with the option rolling out later this year.
The 30-day duration itself is not changing. What moves is the location: Anthropic has spent months working with more than 100 enterprise customers, Salesforce among them, on a system that keeps the required logs inside a customer's own environment rather than Anthropic's.
The Claude Code Precedent Already Shows the Catch
Anthropic tested a version of this idea two weeks earlier. On 6 August 2026 it opened a public beta letting Claude Code sessions run inside a customer's own infrastructure, keeping repository checkouts, build artifacts and secrets on the customer's network rather than Anthropic's.
Organizations that have switched on Anthropic's zero-data-retention setting cannot use those self-hosted environments at all, a direct incompatibility Anthropic states outright. It previews the trade-off arriving with the wider retention change too: a firm can host its own copy of the mandatory logs, but it still cannot make Anthropic stop keeping them.
The Owner's Real Choice
Buying Claude for a regulated workload is no longer just a question of model quality. It is now a decision about where four different categories of data physically land, and the answer differs by product and by account setting.
| Mode | Data location | Retention duration | Availability |
|---|---|---|---|
| Standard cloud retention (Jun 2026) | Anthropic's servers | 30 days, no opt-out | Live now |
| Self-hosted retention (announced Aug 2026) | Customer's own infrastructure | 30 days, same requirement | Rolling out later in 2026 |
| Claude Code self-hosted environments (6 Aug 2026) | Customer network for code, Anthropic API for chat | Not applicable, incompatible with zero-retention | Public beta now |
| OpenAI zero-retention safety system (19 Aug 2026) | No log storage | Zero | Preview now |
Anthropic is not making this move in isolation. OpenAI announced a rival zero-retention safety system just one day earlier, on 19 August 2026, that avoids storing conversation logs at all while still trying to catch misuse. Data-location policy has become a genuine vendor-selection factor for any enterprise choosing an AI provider, not a line buried in the terms of service.
Read next: Microsoft Trained Its Sellers to Call Claude Slower | The AI Store Manager That Fired a Human
