Skip to main content

Comparison

OpenAI vs Anthropic API Data Retention Policies: The Defaults Point One Way, the Fine Print Another

By AInews Desk
OpenAI vs Anthropic API Data Retention Policies: The Defaults Point One Way, the Fine Print Another
6 min read
1133 words

If you send a prompt to Anthropic's API with no special arrangement, the company's documentation says your prompts and Claude's outputs are not retained by default. Send the same prompt to OpenAI's API and abuse monitoring logs are retained for up to 30 days unless you've been approved for something stricter. On defaults, Anthropic looks like the clear winner. The exceptions change that picture. By the end, you'll know which tier your customer data actually falls under.

Training is where the two agree

Neither company trains on API traffic by default. OpenAI says that as of March 1, 2023, data sent to the OpenAI API is not used to train or improve its models, unless you explicitly opt in. Anthropic says retained data is never used for model training without your express permission.

So training doesn't separate them. Retention does.

What OpenAI keeps if you change nothing

OpenAI's standard retention comes from several mechanisms that overlap:

  • Abuse monitoring logs are kept for up to 30 days by default so OpenAI can enforce its usage policies.
  • Application state on the Responses API has a 30-day retention period by default, or whenever store is set to true. Response data in those cases is kept for at least 30 days.
  • Background mode writes response data to disk for roughly 10 minutes so clients can poll for results.
  • Prompt caching can keep encrypted key/value tensors in GPU-local storage, and that data is not retained after the 24-hour expiration.

GPT-Live sessions work differently. Session storage is disabled by default. In projects where it's turned on, store: true keeps the finished recording for 30 days. One thing to know: the API has no public endpoint for deleting stored sessions, so you can't remove a stored session early. If you turn storage on, you live with the full retention window.

OpenAI's Zero Data Retention requires approval

OpenAI doesn't let you switch on Zero Data Retention yourself. Its documentation refers to customers approved for Zero Data Retention and tells buyers to contact sales to ask about eligibility. Once ZDR is on for an organization, the store parameter is always treated as false, even if a request tries to set it to true. That's a sensible safeguard. If a developer leaves store: true in some forgotten code path, nothing gets persisted anyway. The same rule covers GPT-Live sessions, which are eligible for ZDR.

ZDR does come with conditions. Customers who use it are responsible for making sure their users follow OpenAI's policies. OpenAI can also make specific models ineligible for ZDR for specific customers if that's reasonably necessary to investigate or prevent severe risk activity, with advance written notice. When that happens, it may keep and have humans review content its classifiers flag.

Anthropic's default is quieter, but it has big exceptions

Anthropic's starting point is closer to zero. It says it keeps only what is technically necessary for a feature to work, purged on the shortest practical time to live.

The exception is four models. Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, and Claude Mythos 5 are designated Covered Models and require 30-day data retention. ZDR isn't available for any of them unless Anthropic expressly authorizes it. If your organization's retention settings don't meet that requirement, Claude Fable 5 requests come back with a 400 invalid_request_error.

The workaround is to isolate that traffic in its own workspace. A ZDR organization can turn on 30-day retention in a single workspace while the others keep zero retention. Don't use the Default Workspace for this, because it always follows the organization-level setting.

Anthropic's ZDR is set up per organization, through sales

With ZDR in place, Anthropic doesn't store customer prompts or responses at rest after the API response is returned. You request it from the sales team, and it is enabled per organization. Each new organization has to be enabled separately, and enablement doesn't carry over to other organizations under the same account. If your company runs several orgs, check each one. Don't assume coverage carries over.

ZDR also has a narrower reach than the API itself: Claude Free, Pro, Max, and Ultra plans fall outside it, and so does any usage in the Claude Console, including the playground.

If a chat or session is flagged by automated trust and safety systems, Anthropic may retain inputs and outputs for up to 2 years. Some products also fall partly outside ZDR: Claude for Excel is not currently ZDR-eligible, and Claude Code on Claude Enterprise needs its own ZDR offering, available only to qualified accounts and separately enabled. Claude Code used with API keys from a Commercial organization is covered by the API's ZDR.

Two server cabinets side by side in a dim data center aisle, one with dark, empty drive bays and the other with rows of lit storage drives.Two server cabinets side by side in a dim data center aisle, one with dark, empty drive bays and the other with rows of lit storage drives.

How the two compare

DimensionOpenAI APIAnthropic API
Training on API dataNo, unless you opt in (developers.openai.com)No, without express permission (platform.claude.com)
Default prompt/output storageResponses API: 30-day app state (developers.openai.com)Not retained by default (platform.claude.com)
Abuse/safety logsUp to 30 days by default (developers.openai.com)Flagged content up to 2 years (platform.claude.com)
How to get ZDRPrior approval via sales (developers.openai.com)Per organization via sales (platform.claude.com)
Developer error under ZDRstore forced to false (developers.openai.com)No storage at rest after response (platform.claude.com)
Scope of exclusionsSession recording/forking off under ZDR (developers.openai.com)Excel not eligible; Claude Code on Enterprise needs its own ZDR offering (platform.claude.com)

Which one fits your workload

Low retention without a sales call: Anthropic's default on models that aren't Covered Models gets you closest. Prompts and outputs aren't retained, and you don't need to negotiate anything.

Zero retention on frontier models: OpenAI's approved ZDR is easier to reason about. store is forced off, and if a model is ever carved out, you get written notice first. On Anthropic, the Covered Models need 30-day retention unless the company expressly authorizes otherwise, so plan for a separate workspace for that traffic.

Several organizations under one account: Anthropic's per-organization enablement is easy to get wrong. Check each org's status before sending customer data through it.

Either vendor: Plan for safety-driven retention. OpenAI keeps the right to review flagged content on models it carves out of ZDR. Anthropic may keep flagged content for up to two years. Neither vendor's zero-retention offer means data is never stored under any circumstances, so your contract should say exactly which exceptions you're accepting.

Sources

6 min read
•
1133 words

Tags

OpenAI vs Anthropic API data retention policiesAPI data retention zero retention coverageOpenAI API data privacyAnthropic Claude API data retention practices

Related Articles