OpenAI Reaffirms Zero Data Retention and Previews Private Safety Processing for Frontier Model API Customers on August 20

OpenAI on August 20 reiterated its Zero Data Retention (ZDR) option for eligible frontier-model API customers and previewed a new capability called Private Safety Processing (PSP), aimed at letting customers run advanced safety reviews on frontier outputs without losing control of the underlying data. The announcement, reported across industry trackers including SoFarBot and RadarAI on the same day, is the first formal description of PSP since OpenAI began signaling the work to enterprise customers earlier this summer.

What ZDR Actually Guarantees

OpenAI's Zero Data Retention commitment has been the contract backbone of its enterprise offering for more than a year, but the August 20 confirmation is the first time the company has publicly re-stated it alongside the PSP preview. Under ZDR, eligible API customers — those on enterprise agreements that explicitly waive abuse-monitoring retention — have their prompts and responses discarded from OpenAI's servers after the response is returned, with no logging, no human review, and no model-training reuse. The ZDR contract is required for customers operating in regulated industries where prompts may contain personal data, trade secrets, or attorney-client privileged content.

What Private Safety Processing Is

PSP is designed for the gap that ZDR leaves open. Even with ZDR, frontier-model deployments in finance, pharma, and government still need some form of content-safety review — checking for jailbreak attempts, code-execution risks, and outputs that violate policy. PSP runs those checks in an isolated compute environment whose logs are also discarded after the review is complete, with the safety verdict returned to the customer without OpenAI retaining either the prompt or the verdict in its primary systems. SoFarBot, covering the August 20 announcement, described PSP as 「a proposed approach to conducting advanced AI safety work while protecting customer data」 and flagged that the announcement 「leaves several practical questions unanswered,」 including how long PSP isolation lasts per request, whether PSP applies to all frontier models or only a list, and what happens to PSP logs in the rare cases where an investigation is required.

How It Fits the Competitive Landscape

The timing is significant. The week of August 18–20 saw Anthropic, OpenAI's closest competitor, post preliminary Q2 revenue of more than $11.5 billion — a more than 14-fold year-on-year jump that put Anthropic ahead of OpenAI on quarterly revenue for the first time, according to Bloomberg reporting and CNBC coverage on August 15. OpenAI, in parallel, paused training of its next-generation frontier model 「Astra」 over safety concerns, according to Chinese press citing Wall Street Journal reporting on August 18, and launched a dedicated ChatGPT product for teens. PSP and the ZDR restatement together signal a defensive front on enterprise trust: even if Anthropic has the revenue lead, OpenAI is positioning the data-protection surface as a competitive differentiator that is harder to match with a smaller compliance organization.

What Customers Should Do Next

For enterprise buyers evaluating frontier models, the practical checklist is straightforward: confirm ZDR is in the contract for any production deployment that handles regulated data, ask whether PSP is included in the safety-review tier the customer is buying, and request a redlined summary of the retention exceptions. OpenAI said PSP will roll out first to a small set of customers with the most stringent privacy requirements before expanding through the end of 2026. The next material signal will be Anthropic's equivalent — a comparable 「safety review without retention」 capability announced in a way that lets the company's enterprise customers benchmark the two.