What is zero data retention?
A buyer signs off on an AI feature assuming the provider forgets each request the moment it answers. OpenAI keeps abuse logs for 30 days by default. Anthropic keeps API inputs and outputs for 30 days by default. Turning that default off takes an approved agreement, not a setting.
Zero data retention (ZDR)
ZDR
Zero data retention is a deal where an LLM provider does not keep the prompts and answers it handles once it has replied. It swaps the provider's usual retention window, kept for abuse checks and support, for deletion on the spot.
ZDR is not a setting you flip on in a dashboard. OpenAI's API docs describe it as something an account gets approved for. It only covers certain endpoints, not the whole API. OpenAI names chat completions and embeddings as ZDR-ready today. It leaves out endpoints built to hold state, like saved threads, because those exist to keep what you asked them to keep.
Anthropic states a default too. It deletes API inputs and outputs within 30 days of receipt. A zero data retention deal is the named exception, agreed between the customer and Anthropic. It is not a checkbox in the console. Both firms describe the same shape. Keeping data is the default. Zero retention is a deal a buyer has to arrange.
Buyers assume the strictest policy is the default one
A team planning an AI feature over a health record or a legal file asks one question: does the model provider keep our data? The honest answer starts with which product, not which company. A consumer chat plan, an API call and an enterprise deal from the same vendor can each follow a different rule.
OpenAI keeps abuse logs for up to 30 days on API traffic by default. That covers accounts that never asked for anything special. Anthropic's default is the same 30-day window for API inputs and outputs. Neither firm deletes on arrival unless the account holds ZDR approval, and that approval does not stretch across every call a product might make.
This is where a security review goes wrong, quietly. A team reads a provider's trust page, sees zero data retention named, and assumes it covers their account. It covers accounts that asked for it and got it, on the calls that support it. A call built to save state, like a stored chat thread, can sit outside ZDR while a plain answer-only call next to it qualifies.
- Name the endpointChat completion, embedding, assistant thread — each differs.
- Read the defaultDays retained before automatic deletion, stated by the provider.
- Check ZDR eligibilityApproved accounts only, and only on eligible endpoints.
- Confirm in writingA DPA or commercial terms addendum, not a support reply.
- Re-check on upgradeA new feature can call a non-eligible endpoint by accident.
Most reviews stop after the second step.
Zero data retention against the terms it gets mixed up with
| Criterion | What it actually means | The mistake teams make |
|---|---|---|
| Zero data retention | An approved deal where the provider deletes inputs and outputs right after the response, on the calls it covers. | Assumed to be the normal state of any API call. |
| No training on your data | A separate promise that inputs are not used to improve the model. Data can still be kept for a window without being used for training. | Treated as proof the data is not stored at all. |
| Data residency | Where the data sits while it exists. Applies whether retention is 30 days or zero. | Mixed up with retention, though a request can be pinned to a region and still logged for weeks. |
| Encryption in transit | Protects the request on its way to the provider. Says nothing about what happens once it arrives. | Named as if it answered the retention question. |
Common questions
01Does OpenAI store my data by default?
Yes. OpenAI's API docs say abuse logs are kept for up to 30 days on all API use, unless the law asks for longer. Zero data retention removes this, but only for approved accounts on the calls it covers.
02Does Anthropic offer zero data retention?
Anthropic names zero data retention as an exception reached by agreement, separate from its usual default of deleting API inputs and outputs within 30 days. No account gets it without that agreement in place.
03Is zero data retention the same as not training on my data?
No. Not training on customer data is a separate commitment about how inputs are used. Zero data retention is about whether the provider keeps a copy at all. A provider can promise one without the other, so check both clauses.
04Which API endpoints qualify for zero data retention?
It varies by provider and by call, not by account alone. OpenAI names specific ones, such as chat completions and embeddings, while calls built to hold state on your behalf are usually left out. Check the exact list for the calls your build actually makes, not the API as a whole.
