Conversation records still aren't kept
OpenAI said through its official account that it will keep offering Zero Data Retention (ZDR) for frontier models. According to the official announcement, for ZDR-eligible API customers, prompts and model responses aren't retained after a request is processed, and OpenAI employees can't view that content. Unless the customer explicitly opts in, that data isn't used to train models either.
Exceptions required by law remain. Images suspected to be child sexual abuse material (CSAM) are retained for reporting and review even in a ZDR environment.
Catching risk without viewing conversations
The core of this announcement is the new safety mechanism more than the promise to continue. OpenAI explained that "as AI takes on longer and more autonomous work, safety systems also need to look at risk across multiple interactions." Until now, ZDR-compatible safety systems evaluated each request individually, which means some risk patterns slip through in multistep work such as agentic tasks.
Private Safety Processing is a preview feature meant to solve this problem. OpenAI said it's designed to identify recurring or connected risk signals across related interactions without OpenAI employees looking at individual conversation content. The specific detection algorithms and scope of application haven't yet been detailed in the official documentation.
What this means for enterprise adoption decisions
ZDR has been a deciding condition for enterprise customers who need contracts, certifications, and regulatory compliance when adopting the API. This reaffirmation means that condition continues as agentic features expand, which is worth noting for organizations that handle sensitive data and are considering long-running autonomous work features.
However, since Private Safety Processing is still in preview, detailed specs on what data it actually analyzes and how far it reaches haven't been released. If you're considering adopting it, you'll need to check the contract terms and data processing scope separately before general release.
ZDR and no training on your data are different things
ZDR is a promise to not store data after a request is processed, while not training on it is a promise not to use customer content for model training, regardless of storage. OpenAI said it maintains both, but the two clauses are often specified separately in contracts, so the accurate way to confirm how they apply is to check your own API contract terms.