Zero retention: what the term promises, and what it quietly doesn't
Zero retention means the provider stops storing your inputs and outputs after serving the request. It does not mean EU hosting, it does not mean no training, and by default it is off. What the three big API providers actually do, and what to put in the DPA.
Zero retention means the provider does not store your inputs and outputs after it has served the request. That is the entire claim. It is narrower than almost everyone using the phrase believes, and on every major API it is off unless you have applied for it.
This article defines the term precisely, separates it from the two promises it gets confused with, sets out what the three big providers actually do by default, and lists what has to be in the contract for the promise to mean anything.
The definition, and its edges
Zero retention is a commitment that request and response content is processed in memory and not persisted to the provider’s storage once the call completes. No logs of your prompt, no copy of the completion, nothing to retrieve later.
What sits outside that boundary is where the misunderstandings live. Metadata usually survives: account identifiers, timestamps, token counts, error codes. Zero retention is a statement about content, not about the existence of a billing record. A provider can truthfully say it retains nothing while knowing exactly how many requests you made and when.
The second edge is that zero retention describes the provider’s storage, not yours. Your own application almost certainly writes the conversation somewhere: a database, a CRM record, a log line. Zero retention upstream does nothing about that, and in a GDPR audit it is your storage that is asked about first.
The three promises that get conflated
Almost every vendor conversation muddles three separate commitments. They can be held in any combination, and holding one implies nothing about the others.
| Promise | The question it answers | Held by |
|---|---|---|
| Zero retention | Does the provider store the content after the request? | A contract term, usually opt-in |
| No training | May the provider learn from the content? | A separate contract term |
| Data residency | In which region is it processed and stored? | An infrastructure choice |
No training is not zero retention. All three major API providers already commit not to train on business API traffic by default, and all three still retain something for a period, whether the content itself or the abuse monitoring log written over it. The promises are independent: content can be kept for thirty days, reviewed by a human for abuse, and never used for training, all at once.
Zero retention is not residency. A provider can hold nothing at all and still process every request in Virginia. Nothing is stored, but personal data was transferred and processed outside the EU, which is a separate question under Chapter V of the GDPR. This is the confusion that costs the most time in procurement, and it deserves its own treatment: storage residency and processing residency are not the same thing.
What the providers actually do
The default on every major platform is retention, not zero retention. The specifics as of October 2026:
OpenAI generates abuse monitoring logs for all API feature usage and keeps them for up to 30 days by default. Zero Data Retention covers only the endpoints listed as eligible, while Modified Abuse Monitoring excludes customer content from the abuse monitoring logs across all endpoints without that restriction. Both are subject to OpenAI’s prior approval and acceptance of additional requirements, so neither arrives by itself. Since August 2026 there is a third arrangement, Zero Data Retention with Private Safety Processing, which lets OpenAI run automated safety monitoring while the zero retention protections hold. (Data controls in the OpenAI platform)
Anthropic deletes commercial API inputs and outputs within 30 days of receipt or generation. The 7-day default it introduced in September 2025 no longer applies, and the commercial baseline is back at 30 days. (Commercial data retention policy) A Zero Data Retention arrangement is enabled per organization on request through the sales team, and since 9 June 2026 it stops at the models Anthropic designates as Covered Models. Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5 and Claude Mythos 5 require at least 30 days of retention and are not available under zero retention unless Anthropic expressly authorizes it, on the Claude API and on the cloud platforms alike. (API and data retention, Covered Models)
Models sold by Azure, the Microsoft Foundry service that contains the Azure OpenAI models, reviews flagged prompts and completions by automated means by default, and content that passes through that automated review is not stored at all. Human review enters when the automated review falls short of its confidence threshold, and only then do prompts and completions land in an abuse monitoring data store, held in the customer’s own Azure geography. Microsoft’s current documentation names no retention period in days for that store. Customers who meet the Limited Access criteria can apply for Modified Abuse Monitoring, which switches off both the storage and the human review. (Abuse monitoring, Data, privacy, and security)
Default retention: up to 30 days of abuse monitoring logs at OpenAI, 30 days for commercial API content at Anthropic, no documented period at Microsoft. Zero retention on all three is an approved arrangement, and at Anthropic it is withheld from the strongest models.
Provider documentation, checked 7 October 2026
One more edge applies to all of it. Anthropic states that flagged inputs and outputs may be held for up to two years whatever arrangement is in place, zero retention and HIPAA readiness included. A trust and safety flag outranks the contract.
The pattern is consistent and worth naming: zero retention is a commercial arrangement rather than a product feature. If nobody at your company filled in a form, you do not have it, no matter what the marketing page said, and for the newest and most capable models the form may not be on offer at all.
The sub-processor problem
A zero retention promise from your vendor is worth exactly as much as the weakest promise in the chain beneath it.
If an agency builds you an agent, that agency is a processor under Article 28 GDPR and every model provider it calls is a sub-processor. Its own infrastructure may retain nothing while the model provider underneath retains for thirty days, and the sentence “we operate zero retention” would still be technically true, about the wrong layer.
This is the single most productive question in a vendor call: name every sub-processor that sees content, and tell me the retention setting on each. An honest answer is a short list with a policy against each name. A vague answer is the finding.
What to put in the contract
Five things, in writing, in the data processing agreement rather than on a web page:
Every sub-processor named, with the right to be notified of changes. A model provider swapped without notice changes your legal position silently.
The retention period per sub-processor, stated as a number of days. “Minimal” and “as short as technically feasible” are not periods.
Whether human review is possible, and under what circumstances. Abuse monitoring generally implies that a person may read flagged content. That is defensible; discovering it after signing is not.
The training exclusion, contractual rather than editorial. A statement in a blog post is a marketing position. A clause is an obligation.
Deletion timelines including backups. “Deleted immediately” and “removed from backups within 90 days” are frequently both true of the same system. Ask for both numbers.
For the German-language version of this checklist alongside the rest of an AVV, see what a GDPR-compliant AI partner has to bring (auf Deutsch).
Why any of this is worth the effort
Under Article 5(1)(e) GDPR, personal data may be kept in identifiable form only as long as necessary for the purpose. Every day of provider-side retention is a day you have to be able to justify, and the justification has to be yours, not the vendor’s: you are the controller.
There is also the plainer argument. Content that was never stored cannot be produced in litigation, cannot appear in someone else’s breach, and cannot be subpoenaed. Retention is a liability that accrues quietly and only ever becomes visible on the worst day.
None of which makes retention wrong. Thirty days of logs is how abuse gets caught, and abuse monitoring protects you too. The point is that it should be a decision you made rather than a default you inherited.
Frequently asked questions
Does zero retention mean my data never leaves my systems?
No. The request still travels to the provider and is processed there. Zero retention means it is not stored afterwards, which is a different and smaller claim than “nothing leaves the building”.
Is zero retention the same as being GDPR-compliant?
No. It helps with the storage limitation principle in Article 5(1)(e), but lawful basis, information duties, sub-processor contracts and international transfers are all separate requirements that zero retention does not touch.
Do the providers train on our API traffic?
Not by default on business API tiers at the three providers above: that commitment is separate from retention and generally already in place. Consumer products are governed by different terms, which is why company data belongs on the API or business tier rather than in someone’s personal account.
Can we get zero retention on a small contract?
Sometimes, but it is usually gated by plan tier and approval rather than sold on a self-service page. The practical route for a smaller company is to work through a partner who already holds the arrangement, or to choose a deployment where the question does not arise. For individual models the answer can be no at any contract size, as Anthropic’s Covered Models show.
If we self-host a model, is retention solved?
The provider-side question disappears and becomes yours. You now decide what gets logged, for how long, and who can read it, which is genuinely better control and genuinely more work. It is not automatically less retention; most self-hosted stacks log more, not less, on day one.
Sources, all retrieved 7 October 2026: OpenAI data controls documentation, Anthropic API and data retention, Anthropic commercial data retention policy, Anthropic Covered Models, abuse monitoring for Models sold by Azure, Microsoft data, privacy, and security, GDPR Article 5 and Article 28. Provider policies change, sometimes within a quarter, and the point of the article is the questions rather than the numbers.
Strategy Call