Claude Zero Data Retention in the EU

TL;DR

Claude Zero Data Retention is a contract, not a residency switch. After the API response, Anthropic does not keep prompts or outputs at rest. Messages and Token Counting are covered. Batch (29 days), code execution (30 days) and Fable (30 days) are not. In the EU set Bedrock to none, or ask sales for ZDR on the first-party API.

What is Claude Zero Data Retention and what does it cover?

Claude Zero Data Retention means customer prompts and outputs are not stored at rest after the API response returns. The arrangement is per organisation. A new org under the same account does not inherit it. Covered endpoints are Messages and Token Counting, plus Claude Code with API keys from a Commercial organisation. Residency is a separate control.

The API and data retention page, read on 24 October 2026, splits processors. On the Claude API, Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed), Anthropic is the processor and ZDR is a sales request. On Amazon Bedrock and Claude Models on Google Cloud (Google managed), the cloud provider is the processor and you use that platform’s retention controls.

ZDR does not cover the Claude Console including the playground, Managed Agents, Free, Pro and Max, Teams and Enterprise product interfaces, Claude for Excel, or CORS from the browser. Claude Code on Claude Enterprise has a separate ZDR offering. If Claude Code metrics logging is on, usage statistics sit outside ZDR.

HIPAA readiness is the path for PHI, not an add-on to ZDR. It exists on the Claude API and is not available on Claude Platform on AWS or Microsoft Foundry. HIPAA blocks ineligible features with a 400 error. ZDR does not block them: calling Batch or code execution steps outside the arrangement for that data.

Residency and ZDR are two switches. inference_geo accepts only global and us. Workspace geo is us only. How to pin inference to EU member states is in Claude data residency in the EU. The legal matrix of the four routes is in the GDPR comparison of Claude providers.

What does Claude Zero Data Retention cost in the EU?

Claude Zero Data Retention has no list surcharge. You pay the same token rates. The cost sits in features ZDR does not cover, and in the 10 percent premium if you add EU residency on Bedrock or Google. The Batch half-price row is not ZDR-eligible.

Anthropic pricing on 24 October 2026 lists Sonnet 5.5 at 2 USD input and 10 USD output per million tokens. Opus 5.5 is 4 / 20. Haiku 5.5 is 0.10 / 0.50 up to 100,000 prompt tokens and 0.50 / 2.50 above that. Batch halves both sides: Sonnet 5.5 Batch 1 / 5, Haiku 5.5 Batch 0.05 / 0.25. Regional and multi-region endpoints add 10 percent from Sonnet 4.5, Haiku 4.5 and Opus 4.5. US-only inference (inference_geo: "us" or Foundry Data Zone Standard US) is 1.1x. Claude Consumption Units are 0.01 USD; 100 CCU equal 1 USD.

Route	Sonnet 5.5 in / out	Haiku 5.5 in / out (to 100k)	Batch on that row	ZDR status
Claude API, global	2 / 10	0.10 / 0.50	1 / 5 or 0.05 / 0.25	Sales contract; Batch not covered
Bedrock or Google, EU endpoint (+10 %)	2.20 / 11	0.11 / 0.55	50 % of the EU row if Batch is allowed	Bedrock none or Google checklist
US-only, 1.1x	2.20 / 11	0.11 / 0.55	1.1x of the Batch row	Not an EU product
Foundry or Claude Platform on AWS	same USD, billed in CCU	same USD, CCU	Batch only if the feature is allowed	Anthropic, per org or subscription

USD per million tokens, standard, as of 24 October 2026. There is no official euro list. The 10 percent rule and the 1.1x factor stack with cache writes (5-minute 1.25x, 1-hour 2x) and cache reads (0.05x on Sonnet 5.5, 0.10 USD). Prompt caching itself is ZDR-eligible: KV representations live in memory for the TTL. Breakpoints and the EU call sit in Claude prompt caching in the EU. Token rates across clouds are in Claude API pricing for the EU.

Worked example, our assumption, not a vendor quote: 20 million input and 2 million output tokens on Sonnet 5.5 cost 60 USD globally. The same volume on the Europe row costs 66 USD. The same volume as Batch costs 30 USD globally. Under ZDR that Batch discount disappears because the Batch API stores jobs for 29 days. If you need residency and ZDR together, budget 66 USD and drop Batch, not 30 USD.

We recommend ZDR plus an EU endpoint for prompts that hold personal data. For classification with no personal data, residency without a ZDR contract is often enough, as long as Bedrock is none or Google request-response logging stays off.

Which features and models fall outside ZDR?

The eligibility table uses three values: Yes, Yes (qualified), and No. No means stateful. The API does not block those features under ZDR. Covered Models need 30 days regardless of the feature column, unless Anthropic expressly authorises ZDR.

Feature or model	ZDR	HIPAA	What is retained
Messages, Token Counting, adaptive thinking, effort, citations	Yes	Yes	nothing after the response
Prompt caching	Yes	Yes	KV and hashes in memory for the TTL
Structured outputs	Yes (qualified)	Yes	JSON schema for up to 24 hours since last use
Batch /v1/messages/batches	No	No	29 days, async storage
Code execution, programmatic tool calling	No	No	container data up to 30 days
Files API, Agent Skills, MCP connector, Managed Agents	No	No	files, skills, sessions until you delete them
Claude Fable 5, Fable 5.1, Mythos 5, Mythos 5.1	No, unless authorised	30-day requirement	Covered Models, 400 without retention
Flagged content, legal holds	independent	independent	up to 2 years
Activity Feed, local and remote session transcripts	outside	HIPAA: local sessions 30 days	6 years default

Source: API and data retention. CORS is not supported for ZDR organisations. Browser calls must go through a backend proxy.

A ZDR organisation unlocks Covered Models only in a workspace with 30-day retention. The Default Workspace always follows the org. In the Console at platform.claude.com/settings/privacy, pick the workspace, set Standard data retention (30-day retention) and accept the terms. Without that setting, Fable 5 returns 400 invalid_request_error: the organisation or workspace must have data retention enabled. Which model to start with is in Claude Sonnet vs Opus. Fable’s 30-day rule and EU availability are in Claude Fable 5.1 in Europe.

How do you set ZDR on Bedrock in the EU?

On Bedrock, zero data retention is mode none in every Region you call. The setting does not propagate. Frankfurt on none leaves Ireland on inherit, which falls back to the model default. Fable 5 and 5.1 require aws_review and otherwise show as unavailable or return ValidationException.

The Bedrock data retention page orders modes: none then default then aws_review then legacy provider_data_share. none writes no request or response data to durable storage and shares nothing with the provider. On the Responses API, store=true is rejected and background mode is unavailable. store=false alone is not ZDR. Opus 4.8 lists none in allowed_modes: even if the account allows aws_review, that model retains nothing. Fable 5 lists only aws_review and provider_data_share. AWS reviews, Anthropic does not see the content, up to 30 days, and with cross-region inference in the destination Region. Abuse detection repeats the 30-day rule for Fable 5 and 5.1. The Enterprise Frontier Safeguards programme grants ZDR through 31 December 2026, then 30 days again.

The Sonnet 5.5 card lists launch on 28 September 2026, end no sooner than 28 September 2027, 1 million context tokens, 128,000 output tokens, June 2026 cutoff and geo ID eu.anthropic.claude-sonnet-5-5. From Frankfurt the destinations stay Frankfurt, Stockholm, Milan, Spain, Ireland and Paris. The platform around that profile is in Amazon Bedrock explained.

Setup in seven steps:

1. Enable Sonnet 5.5 or Haiku 5.5 in an EU Region. Do not mix Fable into this account while none is mandatory.
2. Set retention to none in every Region you use, control plane and Mantle separately.
3. Read the effective mode. If you see inherit or default, set it again.
4. Pin IAM or an SCP with bedrock:DataRetentionMode or bedrock-mantle:DataRetentionMode to none.
5. Check destinations for eu.anthropic.claude-sonnet-5-5 with get-inference-profile.
6. Send a Messages call to bedrock-runtime in eu-central-1.
7. Keep your own request logs for at least 30 days. Anthropic recommends that for investigations. Those logs are your processing.

# Account retention in Frankfurt, control plane
curl -X PUT https://bedrock.eu-central-1.amazonaws.com/data-retention \
  -H "Authorization: Bearer $AWS_BEARER_TOKEN_BEDROCK" \
  -H "Content-Type: application/json" \
  -d '{ "mode": "none" }'

# Read it back
curl https://bedrock.eu-central-1.amazonaws.com/data-retention \
  -H "Authorization: Bearer $AWS_BEARER_TOKEN_BEDROCK"

# Call on the EU profile
aws bedrock-runtime converse \
  --region eu-central-1 \
  --model-id eu.anthropic.claude-sonnet-5-5 \
  --messages '[{"role":"user","content":[{"text":"State retention mode none in one sentence."}]}]' \
  --inference-config '{"maxTokens":256}'

ZDR access for models that require retention goes through Anthropic, not a console toggle. Approved accounts see "none" in that model’s allowed_modes.

How do Google Cloud and Microsoft Foundry differ?

Google gives a checklist. Foundry gives a subscription switch that Anthropic owns. Neither is a one-for-one copy of Bedrock none. Foundry has no EU data zone. Google Europe excludes the United Kingdom and Switzerland.

The Google zero data retention page, last updated 22 October 2026, lists concrete actions: leave request-response logging off, set store = false on the Interactions API (the default is true), avoid Grounding with Google Search (logs up to 3 days) or use Web Grounding for Enterprise, and avoid Grounding with Google Maps (30 days). Google does not train on customer data without permission. Claude Models on Google Cloud (Anthropic managed) follow the first-party ZDR contract. Claude Models (Google managed) follow Google’s processor documents.

Claude Fable 5.1 on Google Cloud sits under the Advanced AI Safety Addendum. All prompts and responses are stored for up to 30 days. Release 1 September 2026, retirement no sooner than 1 March 2027, 1 million input tokens, 128,000 output, ID claude-fable-5-1. Section F of Anthropic’s Service Specific Terms requires prompt-response sharing with Anthropic per model location. That is not a ZDR path. The residency page states that the Europe multi-region endpoint covers EU member states only.

Microsoft Foundry does not treat retention as an Azure toggle. The troubleshooting page says a Covered Model plus ZDR on the subscription returns 400 invalid_request_error. Anthropic owns the setting. Microsoft cannot change it. New subscriptions have retention on. Either Anthropic disables ZDR for that subscription, or you create a second subscription. The hosting comparison makes Anthropic the processor on both hosting options. Hosted on Azure stores data at rest in the selected Azure geography, with Global or US Data Zone processing. Hosted on Anthropic may run outside Azure. The models page lists only Global Standard and Data Zone Standard (US). Product and CCU are in Azure AI Foundry explained.

We recommend Bedrock none plus eu.anthropic.claude-sonnet-5-5 from Frankfurt when ZDR and EU processing are both mandatory. Use Google Europe when the project already lives there and Fable stays off. Use Foundry only when procurement forces Azure and prompts hold no personal data.

FAQ

Does Claude Zero Data Retention cost extra?

No. The price list has no ZDR surcharge. You pay the normal token rates. The expensive part is dropping Batch (50 percent off, 29-day storage) and adding the 10 percent EU-endpoint premium when you also need residency.

ZDR vs residency: which do we need first?

Residency controls where inference runs. ZDR controls what remains after the response. The Claude API can be on ZDR and still run globally or in the US. If processing must stay in EU member states, pin Bedrock or Google, with or without a ZDR contract. Both together is the conservative pair for personal data in prompts.

Bedrock vs the Anthropic API: which ZDR path?

Bedrock none is self-service per Region. The Anthropic API needs sales, applies per organisation and does not block ineligible features. HIPAA exists only on the Claude API, not on Claude Platform on AWS and not on Foundry. For EU teams with IAM and SCPs, Bedrock is the shorter path.

Does ZDR apply to Claude Code?

With API keys from a Commercial organisation, yes, if that org has ZDR. Metrics logging stays exempt. Claude Code on Claude Enterprise needs its own ZDR offering. Console, Teams UI and consumer plans are not covered.

What happens if we call the Batch API under ZDR?

Nothing blocks the request. The Batch API stores jobs for 29 days. For that data you leave the ZDR arrangement. Under HIPAA you get a 400 error instead.

Is store=false enough on Bedrock or Google?

No. On Bedrock only data_retention_mode: none is guaranteed ZDR. store=false can still trigger safety retention, just without a GET on the response. On Google you must set store = false because the default is true, and you still have to check logging and grounding.

Sources

1. Anthropic: API and data retention (24 October 2026)
2. Anthropic: Pricing (24 October 2026)
3. Anthropic: Data residency (24 October 2026)
4. AWS Bedrock: Data retention (24 October 2026)
5. AWS Bedrock: Abuse detection (24 October 2026)
6. AWS Bedrock: Claude Sonnet 5.5 model card (24 October 2026)
7. Google Cloud: Zero data retention (24 October 2026)
8. Google Cloud: Claude Fable 5.1 (24 October 2026)
9. Google Cloud: Data residency (24 October 2026)
10. Microsoft Learn: Claude models in Foundry (24 October 2026)
11. Microsoft Learn: Deploy Claude in Foundry (24 October 2026)
12. Microsoft Learn: Claude hosting comparison (24 October 2026)