Zero data retention

How it works

OpenAI names two kinds of stored data. Abuse monitoring logs are kept for up to 30 days by default, or longer where the law requires it or to protect its services or others from harm. ZDR removes customer content from those logs on all endpoints, apart from rare image and file cases. The second kind, application state, is what features store to work. ZDR removes it only on eligible endpoints, and even there with limits, such as audio outputs kept for 1 hour. On the Responses and Chat Completions endpoints, store is always treated as false.

Anthropic describes ZDR as not storing “customer prompts or responses at rest after the API response is returned”. Some eligible features keep a small technical record, such as a JSON schema cached for up to 24 hours since last use. ZDR covers eligible API features and Claude Code for Enterprise, not the Console or the Teams and Enterprise chat apps.

On Amazon Bedrock the customer sets ZDR through the API: data_retention_mode: none, under which “No request or response data is written to durable storage”. It is set per Region, and an organization policy can forbid any other mode. Setting store=false on a request is not the same, Bedrock warns.

Why it matters

Retention is one of the data terms to check before a language model provider sees regulated data, as using LLMs securely in regulated industries sets out.

ZDR limits what is kept after the response, not what the provider processes to produce it, and training is a separate promise. For protected health information, Anthropic points organizations to a different arrangement, HIPAA readiness under a signed business associate agreement. Under it, data can be kept with safeguards such as encryption, rather than deleted, and Anthropic says ZDR is not also needed.

Trade-offs

You give up stateful features. On OpenAI, endpoints such as assistants, files, batches and stored conversations keep application state even with ZDR on. On Anthropic, features such as the Batch API, with its 29-day retention, the Files API and code execution fall outside it. Nothing blocks calling one, but doing so steps outside ZDR for that data.

Anthropic’s Covered Models are off the table. Anthropic requires 30-day retention for these models, Claude Fable 5 and 5.1 and Claude Mythos 5 and 5.1, unless it expressly authorizes otherwise. A ZDR organization can turn on retention for a single workspace while its other workspaces keep ZDR. On Bedrock, Fable 5 and 5.1 need the human review mode, which keeps prompts and outputs inside AWS for up to 30 days and lets AWS review them. Bedrock evaluates ZDR access to such models per account and per model, and Anthropic manages eligibility for Claude.

Eligibility can change. OpenAI reserves the right to make models ineligible for specific customers after written notice. Under one of its options, content then goes into encrypted abuse monitoring logs, excluded from human review unless the law requires it. Under another, where needed to investigate or prevent severe risk activity, OpenAI may keep and have humans review content its classifiers flag.

Where definitions disagree

“Zero” does not mean the same thing at each provider. Bedrock’s mode none writes nothing to durable storage, and some models there, such as Claude Opus 4.8, keep no data whatever mode is set. OpenAI’s ZDR removes content from abuse monitoring logs but keeps image inputs flagged as potential CSAM for manual review. Anthropic’s ZDR still keeps data where needed “to comply with law or combat misuse or harm”, keeps flagged content for up to 2 years, and keeps safety classifier results. The contract, not the name, says what is kept.

Questions and answers

What is zero data retention for LLM APIs?

Zero data retention (ZDR) is an arrangement under which a language model provider does not store prompts or outputs after returning the response. OpenAI and Anthropic grant it on approval. Amazon Bedrock offers it as the data retention mode none. On OpenAI and Anthropic, stateful features such as batch jobs keep their own retention even when ZDR is on.

Does zero data retention mean nothing is ever kept?

No. Anthropic may keep flagged inputs and outputs for up to 2 years and keeps safety classifier results; OpenAI keeps image inputs flagged as potential CSAM for manual review. Amazon Bedrock's mode none writes no request or response data to durable storage, but models that require retention are blocked under it.

Is zero data retention the same as not training on my data?

No. Training and retention are separate terms. OpenAI does not train on API data by default whether or not ZDR is on. Without ZDR or its Modified Abuse Monitoring control, OpenAI keeps abuse monitoring logs for up to 30 days, or longer where the law requires it or to protect its services or others from harm.