Is the Claude API GDPR Compliant? It Depends Where You Run It (2026)

The Claude API can be used with personal data in compliance with the GDPR. Anthropic's commercial terms already address several points by default: API data is not used for training, the Data Processing Agreement is part of the Commercial Terms, and inputs and outputs are deleted within thirty days by default.

Those terms cover Anthropic's obligations as processor, whereas the controller's obligations, including the DPIA, data minimisation, the privacy notice and the transfer record, remain with you. This guide accompanies the ChatGPT API guide and follows the same structure: what the provider handles, and what you need to configure and document yourself. For a side-by-side comparison, see GDPR-compliant LLM APIs: OpenAI vs Anthropic vs Google.

Get articles like this. Compliance Engineering, practical AI compliance for engineers and founders, by a CIPP/E certified practitioner.

Need this checked for your setup? A £500 scoping review covers your provider configuration, data processing agreement, data flows, retention and DPIA, with a written report within one week.

The consumer app and the API are separate products

Many GDPR problems with large language models arise from treating the consumer chat product and the API as the same service, although they are governed by different terms.

Consumer Claude (claude.ai Free, Pro and Max):

• Since 8 October 2025, users choose whether their conversations are used for training, with retention of up to five years where they agree
• No Data Processing Agreement
• No data residency or retention controls for business use

Claude API (commercial):

• Inputs and outputs are not used for model training, by default and under the contract
• The DPA is incorporated into the Commercial Terms
• Inputs and outputs are deleted within thirty days by default, and zero data retention is available to qualifying enterprise customers on most models

Where staff enter customer details into the consumer app, the compliance risk lies there, whereas the commercial API, configured correctly, is the appropriate route for business personal data.

What Anthropic's commercial terms provide

Anthropic's commercial terms address training and retention, which reduces the configuration work, although neither point removes your own obligations as controller.

Training. Anthropic does not use inputs or outputs from its commercial products for model training. The position is contractual and applies by default, so there is no per-request setting to manage across a team.

Retention. Anthropic describes its retention in two places, which should be read together rather than merged. Its Privacy Center states that it automatically deletes API inputs and outputs within 30 days of receipt or generation, subject to exceptions, while its developer documentation states that conversation content is not retained by default, except for the covered models described in Step 2. Either way, less data is held by the processor and for a shorter period, which supports data minimisation.

Step 1: The data processing agreement

Article 28 of the GDPR requires a written contract whenever a third party processes personal data on your behalf, and it is likely to be one of the first documents a regulator asks for.

Anthropic's DPA is incorporated into its Commercial Terms of Service, so accepting the Commercial Terms accepts the DPA, and standard deployments need no separate signature. The current version is dated 24 February 2025. For customers in the EEA, Switzerland or the UK, the contracting entity is Anthropic Ireland, Limited and the governing law is Irish law. Disputes go to arbitration in Dublin under the UNCITRAL Arbitration Rules. The DPA incorporates the Standard Contractual Clauses, the UK Addendum and a Swiss addendum for international transfers. Free and Pro plans do not include a DPA, which is a further reason to keep business personal data on the commercial API.

These are the eight clauses to check in a data processing agreement with an AI provider, and a current summary of Anthropic's terms is maintained in the Anthropic compliance review.

Step 2: Retention and zero data retention

Default retention is acceptable for most production systems, provided it is documented, but three conditions apply to the covered models, zero data retention and particular features:

• Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5 and Claude Mythos 5 are designated covered models, whose prompts and outputs are retained for at least 30 days, and Anthropic's Service Specific Terms state that its right to retain and review covered-model data supersedes modified retention commitments such as zero data retention. Anthropic has said that eligible customers will have a time-limited option to use zero data retention with Fable 5 and Fable 5.1, as a transition to its Enterprise Frontier Safeguards programme.
• Zero data retention is available, with Anthropic's approval, for the other models, and is requested through Anthropic's sales team for the whole organisation. Under it, Anthropic does not store prompts or responses at rest after the response is returned, although content flagged by its trust and safety systems may still be kept for up to two years and safety classifier results are retained.
• Several features are outside zero data retention, including batch processing, the Files API, code execution, the Claude Console and the MCP connector, so Anthropic's published list should be checked before you commit to zero retention in a client contract.

Set retention to the shortest period you can justify, because the less data Anthropic holds, the simpler your minimisation and deletion arrangements become.

Where the model runs determines who the processor is

Claude is available through several routes, and the processor differs between them. On Anthropic's own API, and on Claude Platform on AWS and Claude in Microsoft Foundry, Anthropic is the processor and its DPA applies. On Amazon Bedrock and on Google Cloud's Gemini Enterprise Agent Platform (formerly Vertex AI), Anthropic's documentation states that the cloud provider is the data processor, so the relevant DPA is that of AWS or Google. For the covered models the published sources conflict. Anthropic's Service Specific Terms provide that Anthropic processes covered-model data used through a cloud platform under its own DPA. Anthropic's developer documentation, however, states that retained covered-model data on Bedrock and the Agent Platform stays within the cloud provider's environment, and AWS states that model providers have no access to Bedrock prompts and completions. On Google Cloud, customers must also enable sharing of prompts and responses with Anthropic to use Fable 5 and Mythos 5, a requirement Google attributes to Anthropic. Separately, AWS retains all Fable 5 and Fable 5.1 traffic for up to 30 days for abuse detection, with possible human review by AWS. Until Anthropic and the cloud provider confirm the position in writing, your records of processing should name each processor that may apply to the models you use.

Anthropic's own API offers no EU processing option. By default, inference may run in any available geography. Customers can instead require US-only inference, at 1.1 times the standard price. Data at rest is stored in the US. Where processing needs to stay in Europe, the route is a cloud platform, and on Amazon Bedrock it depends on the model. Claude Opus 5 and Sonnet 5 can run in-Region in Ireland (eu-west-1) and Stockholm (eu-north-1), and Sonnet 5 also in London (eu-west-2). In those Regions AWS states that requests never leave the Region you specify. Claude Opus 5.5 and Sonnet 5.5 are offered in the EU Regions through geographic and global routing. Claude Fable 5, Fable 5.1 and Mythos 5.1 are available in the EU Regions only through global routing. A client or DPIA that requires processing to remain in Europe therefore needs a European Region and a model whose routing keeps requests within Europe, as the in-Region models do. This section was checked against Anthropic's data residency documentation and AWS's Regional availability tables on 4 October 2026.

Step 3: Minimise the data you send

Data minimisation is a controller obligation and an area where many integrations fall short, because sending a complete customer record when the model needs only part of it breaches the principle, however favourable the provider's terms.

More than necessary:

"Customer Jane Doe (jane@email.com, account #45678,
DOB 12/04/1990, 42 Oak Street, London) is asking
about her recent order."

Minimised:

"A customer is asking about order status. Question:
'When will my order arrive?' Order date: 12 June.
Expected delivery: 17 June."

Remove names, email addresses, account numbers and other identifiers that the model does not need, and build minimisation into how prompts are constructed. Where personal data must be included, for example in medical, legal or financial queries, record the reason in the DPIA.

Step 4: Carry out a DPIA

A data protection impact assessment records that you assessed the risks before the system went live. For a Claude API integration it should cover:

• The personal data that passes through the API, described specifically, for example "customer first name, support query text and order reference" rather than "customer data".
• The lawful basis, usually legitimate interests or performance of a contract, with the reasons for the choice recorded.
• The risks, including a breach at Anthropic, retention beyond what is expected and the international transfer.
• The safeguards, including the DPA, minimised retention or zero data retention, data minimisation in prompts, encryption in transit, access controls and staff training.

The DPIA guide for AI systems sets out the structure, and a separate guide explains when a DPIA is mandatory. Where the system takes actions as well as producing answers, the DPIA guide for AI agents covers the additional requirements.

Step 5: Update your privacy notice

Tell individuals that you use an AI processor, explaining:

• what the AI does, for example "we use AI to help answer support questions";
• that the processor is Anthropic;
• where the data is processed, which is relevant to transfers;
• how long it is kept;
• their rights of access, erasure and objection.

This information belongs in the privacy notice, and where the AI is used in a chatbot you should also say so at the start of the conversation.

Step 6: International transfers

Anthropic is a US company, so processing EU or UK personal data through its API involves an international transfer. Anthropic has no entry on the EU-US Data Privacy Framework register under any status, which was confirmed again on 2 October 2026, although some vendor summaries state otherwise.

The transfer therefore rests on the Standard Contractual Clauses (Module Two, controller to processor, and Module Three, processor to processor), which Anthropic's DPA incorporates together with the UK Addendum issued by the ICO for UK data and a Swiss addendum for Swiss data. The clauses are governed by Irish law, except that the laws of Switzerland apply to transfers governed by Swiss data protection law, and Anthropic undertakes to provide the information reasonably necessary for you to complete a transfer impact assessment.

Record the transfer mechanism in your records of processing activities as the Standard Contractual Clauses with the applicable addendum, and carry out a transfer risk assessment. Because the transfer does not rely on the Data Privacy Framework, it would not fall away if the Framework were invalidated, as Safe Harbor and Privacy Shield were.

Claude through Microsoft 365 Copilot

If you access Claude through Microsoft 365 Copilot rather than Anthropic's own API, Anthropic's processing falls outside the Microsoft EU Data Boundary. An organisation that has committed to customers that their data will be processed within the EU Data Boundary may breach that commitment by routing requests through the integration. If the EU Data Boundary forms part of your commitments, confirm the route your requests take before relying on it. The Microsoft Copilot guide explains the admin setting that controls this and its effect.

The EU AI Act

Data protection and the EU AI Act are separate regimes. Under the AI Act, Anthropic is the provider of a general-purpose AI model and is subject to the obligations in Articles 51 to 55, which have applied since 2 August 2025. As the provider and deployer of the system you build on Claude, you have obligations of your own, which include the Article 50 transparency duties that have applied since 2 August 2026, and under the Digital Omnibus the high-risk obligations for Annex III systems apply from 2 December 2027. Meeting the GDPR requirements does not satisfy the AI Act obligations, and the reverse is also true.

Certifications and sub-processors

For due diligence, Anthropic's Trust Center lists a SOC 2 Type II report, ISO 27001:2022 and ISO/IEC 42001:2023, the AI management system standard, and Anthropic offers a HIPAA Business Associate Agreement for qualifying healthcare use. Its published sub-processor list has 20 entries, including Amazon Web Services, Google Cloud Platform and Microsoft Azure as cloud infrastructure providers. Anthropic must give reasonable notice of a new sub-processor, and you can object on reasonable data privacy or security grounds within 15 days of that notice. List the sub-processors in your records of processing.

Checklist before going live

Before using the Claude API with personal data in production:

• Business data is processed on the commercial API, not the consumer app
• The Commercial Terms are accepted, so the DPA is in place
• Retention is set to the minimum you can justify, with zero data retention requested for sensitive data
• Data minimisation is built into prompt construction
• The DPIA is completed and recorded
• The privacy notice names Anthropic as a processor
• The lawful basis is identified (legitimate interests or contract)
• The transfer mechanism is recorded (the Standard Contractual Clauses with the UK Addendum and, for Swiss data, the Swiss addendum)
• The hosting route is recorded (Anthropic API, Bedrock or Google Cloud) with each processor named, and, where processing must remain in Europe, a European Region and a model whose routing keeps requests within Europe
• Copilot routing is checked if you rely on the EU Data Boundary
• Sub-processors are listed in your records
• A deletion process is in place on your side

The documents on this checklist, as templates. The £49 AI API GDPR Compliance Pack contains editable Word templates for the contract review, the retention and zero data retention record, the DPIA, the privacy notice wording, the transfer assessment and the records of processing entries. Anthropic's position is recorded as published on 4 October 2026, alongside OpenAI's and Google's, each with its source.

Summary

Anthropic's commercial terms address three points by default, in that API data is not used for training, the DPA is part of the Commercial Terms, and inputs and outputs are deleted within thirty days by default. The remaining work, which is what regulators examine, falls to the controller, who needs to configure the provider correctly and then complete its own minimisation, DPIA, privacy notice and transfer record.

Where AI systems fail in production, the failures can become public and liability can follow. The AI Agent Incident Register analyses real failures of AI agents, including a company held to a policy its chatbot invented.

---

Michael K. Onyekwere is a CIPP/E certified data protection professional at Janus Compliance. For a review of your provider setup, transfers, DPIA and retention, start with a £500 scoping review. If you also need the system built, see the AI Chatbot and Compliance Package.

Anthropic's DPA, retention, residency, covered-model terms, certifications and sub-processors, and the AWS Bedrock Regional availability tables, were re-checked on 4 October 2026, and Anthropic's absence from the Data Privacy Framework register was confirmed the same day. Other terms were re-verified on 17 September 2026 against Anthropic's Data Processing Addendum (dated 24 February 2025), its commercial retention policy (updated 1 July 2026), its platform documentation on retention and covered models, its privacy policy (effective 10 September 2026), and AWS's Bedrock Regional availability and cross-Region inference pages. Other content current as at 18 June 2026. This page is general information, not legal advice. Provider terms change, so check Anthropic's current DPA, retention, hosting and certifications before relying on them. See also: ChatGPT API and GDPR, the GDPR-compliant ChatGPT API setup guide, Is the Gemini API GDPR compliant?, AI vendor due diligence.