Your logs can stay in your own account, for free. Inference still has two settings, and neither is Europe.

What the announcement says

Anthropic says it built EFS with more than a hundred customers across financial services, healthcare, manufacturing, telecoms, law, retail and the public sector. In regulated industries, many of them could not use a model that retains data, and EFS is Anthropic's answer to that.

The package, as the announcement describes it:

• Automated safety monitoring stays in place. It looks for fraud, sophisticated cyberattacks, agents behaving destructively, stolen or leaked credentials, and attempts to build offensive cyber or biological capabilities.
• The monitoring data can be stored in the customer's own cloud account, under the customer's own keys, access policies and audit logging.
• Flags go directly to the customer, who reviews them and decides what to do. No human at Anthropic needs to look.
• It will work the same way on Claude Code, Claude Enterprise and the Claude Platform, and through Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform and Microsoft Foundry.
• Anthropic does not charge for it. Storage, reads, writes and egress are billed by the cloud provider.
• The rollout is phased, with the stated goal of broad availability later this autumn. Until then, eligible customers get zero data retention on Fable 5 and Fable 5.1.

One thing it does not say: where the model runs. No country, no region, no mention of data residency or Europe.

This is an update: part of our 4 September piece is now out of date

On 4 September, answering a Register article from 2 September, we wrote that if you have to apply for data residency, it is a feature, and that you should be able to name the building, not the setting. That Register piece named EFS itself. The vendor's own announcement had been public for three days by then, and we published without reading it. That is on us.

What still stands: Anthropic's documentation still says ZDR is enabled per organisation on request, and that Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1 require 30-day retention unless Anthropic expressly authorises otherwise. The application step has not gone away.

What is out of date: our storage argument. Under EFS the monitoring data lives in the customer's own account, under the customer's own keys. That is far closer to a property than to a setting. If your requirement was that the logs sit in your storage and nobody else reads them, Anthropic will give you that for free from this autumn. That is a sentence we have to sign our name to.

Storage location is now a contract term. Processing location is not.

What remains is a different question. Storage location says where the data rests after the model is done. Processing location says where the model runs while it works on your prompt. EFS settles the first. The announcement is silent on the second, so we went to Anthropic's own data residency documentation.

It describes two controls. The inference_geo parameter decides where inference runs, and it takes two values: global, meaning wherever capacity is available, and us, meaning US infrastructure only. The workspace geo, which covers data at rest and auxiliary processing, has exactly one value today: us. No European value appears on either list, and the documentation's own limitations section says so. US-only inference on Claude 4.6 and later is billed at 1.1x across every token category.

This matters because contracts ask about processing too. Data processing agreements normally fix where processing happens, not only where data is stored. Financial and healthcare supervisors ask the same question about outsourcing, and AI Act documentation does not skip it either. If your internal policy says personal data is processed inside the EU, EFS leaves you exactly where you were: you can pick one of two values for where the model runs, and neither is Europe.

What we think, including the parts that cut against us

This is the strongest source against us this week, so the part against us comes first.

First: EFS is free. No capex, no GPU purchase, no operations team. For a company with no hard requirement on processing location, only a requirement that the logs stay at home and nobody reads them, EFS is simply a better answer than ours. It is cheaper, faster to adopt, and it comes with Anthropic's top models.

Second: for many buyers, storage location is the whole question. To them, the distinction between storage and processing will sound like hair-splitting. We cannot tell you in advance which word your contract uses, only that it is worth checking.

Third: EFS has not shipped. We do not judge a product that neither we nor our customers have used; everything above is about the announcement and the documentation.

Fourth: the price list is not on our side this week either. Anthropic's own pricing page says the Sonnet 5 increase to $3 and $15 per million input and output tokens, scheduled for 1 September, will not happen; $2 and $10 is now the standard price, and Fable 5.1 cache hits cost $0.25 per million tokens. If you have read us arguing that public AI only gets more expensive, you should see those numbers too.

Fifth: we have an interest. We sell the thing we say EFS does not cover, which is why we set out what it does cover first.

What we do instead, and what you give up for it

With us, processing location is fixed, because there is no other route. Private ChatGPT runs open-weight models on private EU GPUs behind your company login. Prompts, documents, logs and embeddings stay on the platform, on EU infrastructure, with no US data processor in the path. The data does not leave the company. ATAILA Cloud and ATAILA Enterprise differ in whose hardware it runs on; the data path is the same in both. The GDPR and CLOUD Act side is covered in our GDPR-compliant AI guide.

Where we stand

We have no quarrel with the announcement. Your storage, your keys, your logs, no human review, no charge. It answers a real request from regulated customers. We withdraw the storage part of our 4 September article.

Our conclusion is narrower than it was two days ago. Storage location can be solved by contract today. Processing location cannot: the vendor's own documentation offers two values, and neither is Europe. If your policy only speaks about storage, EFS is enough, and it is cheaper than us. If it speaks about processing, the model has to come closer, and you need to know what you give up for that.

One testable sentence: open your data processing agreement and check whether the word next to the location is "storage" or "processing", because today those two words have two different answers.