In short: anyone using Claude Code on a project that holds real customer data needs commercial access, either directly from Anthropic or through a cloud platform such as Amazon Bedrock. The Pro, Max and Ultra subscriptions come with no data processing agreement, however they are configured. Processing in the EU is possible, but not through a region choice at Anthropic.
Two contract worlds
Anthropic separates by account: Free, Pro, Max and Ultra run under the consumer terms, Team, Enterprise and the API under the commercial terms. Claude Code itself is the same either way.
Only the commercial terms contain a data processing agreement (DPA) with the EU Standard Contractual Clauses. If you process personal data on behalf of a customer, Article 28(3) GDPR requires exactly that agreement. A DPA on its own does not make a use case lawful – but without it the basis is missing.
The training switch is not enough
For consumer accounts, one setting decides whether Anthropic may use sessions for training, in Claude Code as well. With training, the data is kept for five years; without it, thirty days. Switching it off only applies from that moment on. Conversations flagged by the safety filters may still be used by Anthropic for its internal safety models.
Also worth knowing: Claude Code stores session transcripts in plaintext under ~/.claude/projects/ for thirty days. With a subscription, usage metrics and error reports are on by default, /feedback sends the transcript including code to Anthropic, and before every WebFetch the hostname goes to Anthropic. This is documented and mostly switchable off, but it belongs in the project notes.
Option 1: subscription
A fixed monthly price, but no DPA. As long as the project holds no personal data of third parties and no customer trade secrets, that is fine. Otherwise it is not.
Option 2: Team, Enterprise or API
A DPA with Standard Contractual Clauses, no training without consent, thirty days of retention. Zero data retention only comes with a separate approval.
The location stays open. The API knows only global (the default) and us; there is no EU option. Anyone who has to name the location in a transfer impact assessment sets us and, from Claude 4.6 onwards, pays a factor of 1.1 on all token prices. The "10 percent for Europe" in some articles probably comes from there; it applies to the United States. Processing there stays a transfer to a third country, and assessing the level of protection is on you.
Option 3: Claude through Amazon Bedrock
Processing in the EU, a DPA with AWS instead of Anthropic, your own keys and logging. Billed per token.
Careful with the region: over the standard endpoint bedrock-runtime, current Claude models offer no In-Region routing. The EU profile (eu.anthropic.…) spreads requests across six EU regions, Frankfurt to Paris. That is "in the EU", but not "in Frankfurt". A single region is only available over the newer endpoint bedrock-mantle, and which one that is depends on the model. Check it before you write a region into an offer.
AWS does not state an explicit EU surcharge. What can become expensive is consumption: an agent that reads files and calls tools needs many times what a chatbot needs. You have to estimate the monthly cost in advance.
Option 4: an open model in your own house
No DPA needed, no third country, the data stays in your own server room. But: on the hardest specialist questions cloud models are better, and the hardware costs money once. For correspondence, summaries and comparisons a local model is enough. The AI Act applies here too: oversight, documentation and training remain mandatory. The stages and indicative prices are under our services.
Comparison
| Route | DPA | Where processing happens | Billing |
|---|---|---|---|
| Subscription: Pro, Max, Ultra | no | worldwide | monthly price |
| Team, Enterprise, API | yes | worldwide or US (factor 1.1) | per token |
| Amazon Bedrock | yes, with AWS | EU profile: six regions; single region over bedrock-mantle |
per token |
| Open model in your own house | not needed | own server room | hardware, then electricity |
What does this mean for your project?
Data first, then the tool. Often test data is enough for development, and the real data only runs over the local server at the customer later. That is cheaper data protection.
If it should be the cloud after all, three pieces of information belong in the project files: which contract applies, where the processing happens, and how that appears in your customer's record of processing activities. A cheap subscription does not do that.
Sources
- Anthropic, data usage in Claude Code: code.claude.com/docs/en/data-usage (external link)
- Anthropic, consumer terms (external link) and commercial terms (external link)
- Anthropic, when the DPA applies (external link)
- Anthropic, model improvement settings (external link)
- Anthropic, API data residency (external link) and server locations (external link)
- Amazon Bedrock, regions per model (external link), model card Claude Sonnet 5 (external link), pricing (external link)
Not legal advice
What we describe here is only what we found in the providers' own documentation. Whether a deployment is lawful in your company is something you settle with your legal department or your lawyer.
