Anthropic's zero data retention arrangement carries 9 documented exclusions, among them Console usage, Claude Managed Agents, Claude for Excel, Covered Models and flagged content. On Amazon Bedrock and Google Cloud the arrangement does not apply at all, because the cloud provider is the data processor there. Pexon reviews the contract and the retention configuration against one real workload.
The starting point
Anthropic's data processing agreement is automatically incorporated into its commercial terms, so a company accepts it by accepting the terms. There is no negotiation, no redline and no signed PDF for the contract file — which is why, when the data protection officer asks for the DPA and its effective date, the answer usually arrives as a public URL and a shrug. Zero data retention is then produced as the reassurance that closes the conversation. It is a real arrangement and it is worth having, but it is narrower than the phrase, and the documentation that says so is not the documentation anyone reads.
The evidence
| What Anthropic holds | Stated retention | Under a zero data retention arrangement |
|---|---|---|
| Inputs and outputs on the Claude API under standard commercial terms | Deleted within 30 days of receipt or generation | Not stored at rest after the API response is returned |
| Inputs and outputs sent to Claude Fable 5 or Claude Mythos 5 | 30 days, required since 9 June 2026 | Not available — Covered Models cannot run under zero data retention |
| Inputs and outputs flagged as a Usage Policy violation | Up to 2 years | Retained regardless of the arrangement |
| Trust and safety classification scores | Up to 7 years under the standard commercial policy | User Safety classifier results are still retained; no period is published for that case |
| Compliance API activity feed and remote session transcripts | 6 years | Outside the arrangement — the Compliance API has its own retention model |
Five retention clocks run at the same time, and a zero data retention arrangement stops one of them. The rest are documented in three different places, none of which is the DPA. This is the table to put in front of whoever is about to write “no data is stored” in an internal memo.
Read on 4 August 2026 from Anthropic’s API and data retention documentation, which lists the nine exclusions, the six-year Compliance API retention and the two-year figure for flagged content, and from Anthropic’s commercial data retention policy, which states the 30-day default and the seven-year figure for trust and safety classification scores. The contractual figures come from the Data Processing Addendum effective 24 February 2025, whose subprocessor list is published at anthropic.com/subprocessors. All four pages change without a version history you can subscribe to. Re-read them before this table is quoted in a contract, an assessment or a supervisory submission.
What we build
01
Which Anthropic terms your organisations are on, when each accepted them, and what that pulls in: Standard Contractual Clauses under Module Two or Module Three, fifteen days to object to a new subprocessor, forty-eight hours for a security breach notification, thirty days for deletion or return after termination. Four numbers, each with an owner and a place to live in your record of processing.
02
Zero data retention is enabled per organisation, not per account, and a workspace can carry a 30-day retention override so it can reach Covered Models while the rest of the organisation stays at zero. We enumerate what is actually set, per organisation and per workspace, and reconcile it against what the rollout deck claims is set.
03
On Amazon Bedrock and Google Cloud the cloud provider is the data processor, and the zero data retention arrangement Anthropic describes for its own API does not reach either platform. Those are also the only two platforms that can keep Claude inference in the EU, so the residency answer and the retention answer come from two different contracts and have to be collected separately.
04
The nine documented exclusions turned into a list of the ones your estate touches: the Console your team pastes into, the Managed Agents session transcripts that persist until deleted, Claude for Excel in finance, the consumer accounts nobody decommissioned. Each becomes a line in the record of processing rather than a surprise in an audit.
05
Some answers are not published anywhere. How long User Safety classifier results survive under your zero data retention arrangement is the clearest example: seven years is documented for classification scores under the standard commercial policy, and nothing is documented for the ZDR case. We write those questions in a form your account team can answer in writing, which is the form a supervisory authority accepts.
The artefact
This is the list we work through in the blueprint. Each line is answered by a document, a console setting or an email from your account team — none of them is answered by a vendor assurance in a slide. Send it to whoever is proposing the rollout and see how many come back.
Line four does not need a meeting. Anthropic documents that a request to a Covered Model from an organisation whose retention configuration does not meet the 30-day requirement is rejected, so the rejection itself reads the setting. Run this once per API key, because the setting is per workspace and the key is what carries it:
curl -s -w '\nHTTP %{http_code}\n' https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{"model":"claude-fable-5","max_tokens":1,
"messages":[{"role":"user","content":"ping"}]}'Read the message, not only the status. HTTP 400 with an invalid_request_error saying “In order to access this model, your organization or workspace must have data retention enabled” means that key’s workspace is at zero data retention. HTTP 200 means 30-day retention is switched on for that workspace and one token has just been spent proving it. A 401 or a different 400 message is an authentication or request problem and says nothing about retention.
No. Anthropic states that the DPA, including Standard Contractual Clauses, is automatically incorporated into the Commercial Terms of Service, so accepting the terms accepts the DPA. The practical consequence is that there is no signature page to file: the date your organisation accepted the commercial terms is the date the DPA took effect, and that date has to be recorded somewhere other than an email thread.
Under a zero data retention arrangement, Anthropic does not store customer prompts or responses at rest after the API response is returned. It is enabled per organisation rather than per account, and Anthropic states that each new organisation needs it enabled separately. A company that creates a second organisation for a new business unit does not inherit the arrangement.
Anthropic documents nine exclusions: Claude Console and the Workbench, Claude Managed Agents, the consumer Free, Pro and Max plans, the Claude Teams and Claude Enterprise product interfaces, Claude for Excel, the Covered Models Claude Fable 5 and Claude Mythos 5, third-party integrations, browser requests using cross-origin resource sharing, and flagged content or legal holds. Most enterprise estates touch at least three of them.
No. Anthropic's documentation states that on Amazon Bedrock and Google Cloud the cloud provider is the data processor, and that the zero data retention arrangement described for the Claude API does not apply there. This matters because those two platforms are also the only two that can keep Claude inference inside the EU, so the residency answer and the retention answer come from different contracts.
Anthropic's commercial retention policy states that for Usage Policy violations it retains inputs and outputs for up to 2 years and trust and safety classification scores for up to 7 years. Anthropic also states that User Safety classifier results are still retained under a zero data retention arrangement, without publishing the period that then applies. That gap is worth a written question to your account team.
Anthropic designates Claude Fable 5 and Claude Mythos 5 as Covered Models and requires 30-day retention of prompts and outputs for them, a rule that took effect on 9 June 2026. Zero data retention is therefore not available for either model. A workspace can have the 30-day retention switched on while the rest of the organisation stays at zero, which is the usual way to keep the exception contained.
Fifteen days. Anthropic's DPA gives the customer fifteen days from the notice to object in writing, and states that a customer who does not object within that window is deemed to have consented. Fifteen days is a review cycle, not a policy, so it needs a named person and a recurring calendar entry rather than a paragraph in a governance document.
Anthropic's DPA commits to deleting or returning customer personal data within thirty days of termination or expiration of the agreement. Separately, the DPA requires notification of a security breach in writing without undue delay and in any event within 48 hours of Anthropic becoming aware of it. Both numbers belong in your incident response plan, not only in the contract file.
A two-week Readiness Blueprint, fixed at €4,900. We read the Anthropic terms that actually apply to your organisations, test the retention configuration of each workspace against one real workload, and hand back the scope statement your record of processing needs plus the list of questions only your account team can answer. All prices are net and exclude VAT.
Next step
Two weeks, fixed at €4,900. We read the Anthropic terms that apply to your organisations, enumerate the retention configuration of every workspace, test it against one real workload, and hand back the scope statement, the checklist below with your answers filled in, and the questions that still need your account team. The plan is yours to keep, whoever implements it. All prices are net and exclude VAT.