Anthropic's data-retention reversal: what it means for mid-sized firms
Anthropic walked back its 30-day retention rule for its top models after 12 weeks. For most 200-person firms the terms never applied. The lesson does.
| What changed first | from 9 June 2026 Anthropic required at least 30 days of retention for all traffic on its Covered Models, Claude Fable 5 and Mythos 5, including for customers with zero-data-retention agreements |
|---|---|
| What changed back | on 1 September 2026 Anthropic announced Enterprise Frontier Safeguards, which keeps retained monitoring data in the customer's own AWS, Azure or Google Cloud storage under the customer's own keys |
| Time between the two | 84 days, or 12 weeks, our arithmetic from Anthropic's own dates |
| Who reviews flagged activity under the new scheme | the customer's own team; Anthropic says no Anthropic human review is required |
| Price | Anthropic charges nothing for Enterprise Frontier Safeguards; the cloud provider bills for storage and data operations |
| Who was consulted | more than 100 customers, Anthropic says, across finance, healthcare, manufacturing, telecom, law, retail and the public sector, including Comcast, KPMG, Mastercard, Salesforce and Visa |
| Who the June rule touched | only organisations that had zero data retention set up; Anthropic says consumer plans (Free, Pro, Max) were unaffected because their data was already retained |
Anthropic reversed course on data retention on 1 September 2026, 12 weeks after it began requiring at least 30 days of retention on all traffic to its most capable models, Claude Fable 5 and Mythos 5. Its replacement, Enterprise Frontier Safeguards, lets large customers keep that data in their own cloud and review flagged activity themselves. For most companies of around 200 people, neither the rule nor the fix changes anything they had.
The verdict: this is not a reason to switch AI vendors, and it is not a scheme you can use. It is a reason to write down, this month, what your staff may and may not put into an AI tool, and to check your vendor's terms on a calendar rather than on trust.
What actually happened
On 9 June 2026, alongside the launch of Fable 5 and Mythos 5, Anthropic designated them Covered Models. Prompts and outputs on those models are kept for at least 30 days so that automated systems can look for misuse, including attacks spread across many sessions and accounts. The part that caused the trouble was narrow: customers who had negotiated zero data retention, meaning Anthropic stores nothing, could no longer use the new models on those terms.
Anthropic said it would use the retained data only for safety, not for training. Its biggest customers pushed back anyway. By Anthropic's own account it spent hundreds of hours with more than 100 of them, from banks to Comcast, KPMG, Mastercard, Salesforce and Visa, and on 1 September announced the alternative.
What does the new scheme ask of a customer?
Under Enterprise Frontier Safeguards, the retained data sits in the customer's own storage on AWS, Azure or Google Cloud, encrypted with keys the customer controls. Anthropic's automated monitoring still runs, but when it spots a pattern, the signal goes to the customer, and the customer's people decide what it means. Anthropic charges nothing for it; the cloud provider bills for the storage. It rolls out in phases from this autumn.
That is a good deal for a bank with a security operations team. It is a job description for everyone else. Someone has to own the storage account, manage the keys, and read and act on alerts about possible misuse by your own staff. The Register's summary of the scheme was blunt about the trade: the retention problem goes away, and the checking becomes yours.
The fix moves the data to your cloud and the alerts to your desk. Without a security team, that desk is empty.
Does this apply to a 200-person company?
Almost certainly not. Anthropic's zero-retention agreements are approved organisation by organisation through its sales team and cover API and certain enterprise Claude Code use; its privacy centre says the June change applied only to organisations that had zero retention set up, and that consumer plans were unaffected because their data was already being kept. If your staff use a normal business plan, your data was already kept under the standard terms, and those did not move.
The standard terms you are actually on
Anthropic's commercial retention page, dated 1 July 2026, says API inputs and outputs are deleted within 30 days, and that business-plan chats a user deletes are purged from its systems within 30 days. Content its automated systems flag as breaking the usage policy can be kept for up to two years, and the safety scores attached to it for up to seven. Other AI vendors publish their own version of this page, and it is the one worth reading before any other.
What it does not mean
It does not mean Anthropic is careless with data, or that a rival is safer. It does not mean you need a zero-retention contract; for most general office work, standard business terms are adequate. And it does not mean the matter is settled. CNBC reported that the June policy still applies to Anthropic's non-enterprise subscribers on the Mythos-class models.
What it does show is that a vendor's retention terms can change on a stated date for customers already using the product, with the largest customers the only ones able to negotiate them back.
What to do this month
Write a one-page rule for your staff with three columns: what may go into an AI tool freely, what may go in only on the company's business plan, and what never goes in at all. Customer personal data, pricing, unreleased drawings and anything under a confidentiality clause belong in the last column until someone has read the terms. Then put a quarterly reminder in one person's diary to open your vendor's retention page, note its date, and flag any change. That costs an afternoon, and it is the part of this story a 200-person company can actually use.