Guides › How to get zero data retention

How to Get Zero Data Retention From OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock and Google Cloud

Zero data retention is requested, not switched on. OpenAI and Anthropic grant it per organization after approval through their sales teams. Azure OpenAI reaches the equivalent through a Limited Access application for modified abuse monitoring. Amazon Bedrock exposes it as a data retention mode you set per Region, with approval needed only for models that require retention. Google Cloud has no single switch and publishes a list of features to turn off instead. This page covers who to ask on each platform, where the setting lives once it is granted, how to confirm it, what you give up, and what to do if the answer is no.

Getting ZDR in four points
  • Ask the party you contract with. ZDR on the direct OpenAI or Anthropic API does not carry over to the same model on a cloud platform. Each has its own approval.
  • Get the scope in writing. Organization, endpoints, models and exclusions. Each new organization or project needs its own enablement.
  • Verify in the account. OpenAI shows a Data Retention tab. Anthropic shows the retention period in Console privacy settings. Azure shows ContentLogging set to false. Bedrock returns the retention mode from its API.
  • Rebuild around it. Stored responses, files, batch, hosted vector stores and code execution sit outside ZDR on every provider.

Provider by Provider

The table paraphrases each provider's documentation as of November 2026. The ZDR explainer covers what the commitment means. This page covers how to obtain it.

ProviderHow you get itWhere it is setHow to confirm
OpenAI APIPrior approval by OpenAI and acceptance of additional requirements. Contact sales.Settings, Organization, Data controls, Data Retention tab. Choose Zero Data Retention or Modified Abuse Monitoring for the organization, with per-project overrides.The tab appears only after approval. Check each project's setting and the endpoint eligibility table.
Anthropic APIContact sales. Requests are reviewed and enabled per organization, and each new organization is enabled separately.Applied by Anthropic to the organization. A single workspace can be moved to 30-day retention to use Covered Models.Claude Console privacy settings show the data retention period. The contract states the scope.
Azure OpenAIApply for modified abuse monitoring through Microsoft's Limited Access process.Applied by Microsoft to the approved resources.The resource's capabilities list shows ContentLogging set to false in the Azure portal or CLI.
Amazon BedrockSelf-service for models that allow it. Approval through your AWS account team for models that require retention.Data retention mode set to none, per Region, at the account or project level through the API.Read the mode back from the data retention API. Enforce it with an IAM policy or service control policy.
Google CloudNo single grant. Follow Google's zero data retention list. Partner models such as OpenAI flagship models need approval per project.Per project: cache configuration, request-response logging and the store setting on the Interactions API.Read the project's cache configuration and logging settings.

Seven Steps That Apply on Every Provider

  1. List the workloads that need it.ZDR is granted to an organization, project or account, so decide which applications and data categories require it before you ask.
  2. Check feature and model eligibility first.Walk each integration for files, batch, stored conversations, hosted vector stores, code execution and connectors, and for models that require retention.
  3. Make the request through the contracting party.OpenAI or Anthropic sales for the direct APIs, Microsoft's Limited Access form for Azure OpenAI, your AWS account team for restricted Bedrock models, your Google account team for partner models.
  4. Get the scope in writing.Organization or account identifiers, endpoints, models and stated exclusions, in an order form or amendment.
  5. Configure and confirm.Set the control at the organization, project or Region level and read it back from the account.
  6. Bring your own systems into line.Gateways, tracing, evaluation tools and vector stores hold the same content. Set their retention to match.
  7. Re-check when models or features change.New models can carry retention requirements, and new features are often stateful.

What Approval Involves

None of the providers publishes approval criteria or a timeline. What they do publish is the responsibility that comes with approval. OpenAI states that customers who enable ZDR or Modified Abuse Monitoring are responsible for ensuring their users follow its usage policies and for meeting moderation and reporting obligations under applicable law, because the provider's own logs will no longer contain the content. Microsoft makes the same point about modified abuse monitoring: detection may be less accurate without human review, and customers should be prepared to respond to abuse notifications. Expect to describe the use case, the data involved and the controls you run yourself.

OpenAI offers approved customers two related controls. Modified Abuse Monitoring excludes customer content from abuse logs across all endpoints and leaves platform features intact. Zero Data Retention does the same and also changes endpoint behavior by forcing stored responses off. A newer variant, ZDR with Private Safety Processing, lets OpenAI run automated safety monitoring while the customer supplies the storage, and carries additional technical requirements.

What You Give Up

ZDR removes features that depend on stored state. On OpenAI, the store parameter on the Responses and Chat Completions APIs is always treated as false, and the Conversations API, Assistants threads, files, vector stores and batch remain outside the arrangement. On Anthropic, an organization with ZDR cannot call the API directly from a browser, and batch, the Files API, code execution, the MCP connector and Managed Agents are outside it. Anthropic's Covered Models require 30-day retention; a ZDR organization can enable that retention for one workspace to use them while its other workspaces stay at zero. On Amazon Bedrock, setting the mode to none blocks any model that requires retention, and the request returns an error.

Most of the excluded features are the ones retrieval systems use. The page on ZDR for RAG and knowledge bases covers how to build around them.

The Cloud Platform Route

A model reached through a cloud platform is governed by that platform's terms. Anthropic states that on Amazon Bedrock and on Google-managed Claude deployments the cloud provider is the data processor, while on Claude Platform on AWS and Anthropic-managed Claude on Google Cloud its own retention policy applies and ZDR is requested from Anthropic. Google states that ZDR approvals for the direct OpenAI API do not apply to OpenAI models on its platform, where ZDR is off by default and approved per project.

On Bedrock, OpenAI models other than the open-weight ones have classifier-flagged traffic retained for up to 30 days unless AWS grants full ZDR. The Claude Fable 5 family requires 30-day retention with possible human review by AWS, unless the customer qualifies for the Enterprise Frontier Safeguards program. The cloud platform comparison sets the three platforms side by side.

If the Answer Is No

Organizations that do not qualify, or that depend on excluded features, have four options. Modified Abuse Monitoring on OpenAI keeps platform features while removing content from abuse logs. For protected health information, Anthropic describes HIPAA readiness under a business associate agreement as the arrangement to use and says ZDR is not also required. A cloud platform may offer a retention mode the direct API does not. And for data that should not be disclosed at all, the answer is to keep it from crossing: substitute identifiers before the request is sent, or run the model in an environment you control, as the sovereign deployment guide describes.

ZDR is also an API arrangement. If the system in question is ChatGPT Enterprise, Claude Enterprise, Microsoft 365 Copilot or Gemini, the page on ZDR on enterprise AI platforms covers what those products offer instead.

Where Hardshell Sits

Hardshell reduces how much depends on the outcome of a ZDR request. It connects enterprise data to ChatGPT, Claude, Microsoft 365 Copilot, Gemini and model APIs without migrating it, enforces each user's existing permissions at query time, substitutes stand-ins for real identifiers before content reaches the provider, and records every retrieval under opaque IDs. With ZDR granted, the provider keeps nothing and never saw the identifiers. Without it, what the provider retains is content with the identifiers already removed. Hardshell deploys as a cloud service, in a private cloud or as an offline bundle.

Frequently Asked Questions

Can I turn on zero data retention myself?

On most providers, no. OpenAI and Anthropic require approval before the control exists in your account, and Azure OpenAI requires an approved Limited Access application. Amazon Bedrock is the exception: you can set the data retention mode to none yourself for models that do not require retention.

Who qualifies for zero data retention?

The providers do not publish criteria. OpenAI describes it as available to eligible customers subject to prior approval and additional requirements, and Anthropic reviews requests per organization through sales. Microsoft frames modified abuse monitoring around highly sensitive or confidential data and customers who cannot permit human review of their prompts.

Does ZDR apply to every API key in my company?

No. It applies to the organization, project or account that was approved. Anthropic enables each organization separately, and OpenAI lets you set the control per project. A team using a key from a different organization is not covered.

Does ZDR on the OpenAI or Anthropic API carry over to Bedrock, Azure or Google Cloud?

No. Each platform has its own terms and its own approval. Google states that ZDR approvals for the direct OpenAI API do not apply on its platform, and Anthropic states that on Amazon Bedrock the cloud provider is the data processor.

What is the difference between ZDR and Modified Abuse Monitoring?

Both exclude customer content from OpenAI's abuse monitoring logs. Modified Abuse Monitoring leaves platform features unchanged. Zero Data Retention additionally forces stored responses off, and endpoints that keep application state remain outside it.

How do I prove zero data retention to an auditor or customer?

Use three pieces of evidence: the written confirmation or contract term naming the organization and scope, the account setting read from the provider's console or API, and an inventory showing your integration calls only eligible endpoints and models. A SOC 2 report describes the provider's controls in general and does not attest to your organization's setting.