Anthropic and OpenAI both default to retaining user prompts and model responses for up to 30 days to support safety review, with narrow enterprise exceptions.
Anthropic announced in July 2026 that it would apply a 30-day data retention period to user prompts and model responses for its top-tier AI models, such as Mythos, in order to run safety checks confirming that AI use does not involve security threats, fraud, or abuseCITE:E3. OpenAI's standard API similarly defaults to caching user prompts and responses on its backend servers for up to 30 daysCITE:E4. Healthcare and financial enterprises can apply for review to enable zero data retention (ZDR); once approved, their data stops being collected starting from the next interactionCITE:E5.
OpenAI previewed Private Safety Processing on August 19, 2026, to reduce enterprise customers' data privacy concerns about its AI model platformCITE:E1.
Under the existing ZDR safety system, each individual interaction between a user and the AI is reviewed separately; Private Safety Processing instead extends review across multiple related interactions, letting an automated system identify overall patterns without any OpenAI personnel accessing enterprise customer contentCITE:E6. OpenAI is also developing a storage option in which customer content is encrypted using keys the customer controls, with OpenAI holding no key backup and therefore unable to access the contentCITE:E7.
OpenAI receives a clearly defined signal pointing to specific activity when its system detects risk, without granting staff access to the underlying contentCITE:E8.
That signal can be used to decide whether the activity needs to be reported to law enforcement; the customer who receives it can first investigate the alert and determine whether reporting is warrantedCITE:E8. Even when activity is flagged this way, OpenAI personnel still cannot access the customer's contentCITE:E8.
Coverage of the announcement frames Private Safety Processing as a way for OpenAI to differentiate itself from Anthropic's data retention approachCITE:E2.
TechCrunch and Axios both said OpenAI's move is intended to give enterprises a stronger incentive to adopt its platformCITE:E10. The Wall Street Journal reported this week that Anthropic's revenue grew to $11.6 billion, nearly doubling from the prior quarter and surpassing OpenAI's revenue for the first time; OpenAI's revenue for the same quarter was $6.7 billion, up 18% quarter over quarterCITE:E11.
| Metric | Anthropic | OpenAI |
|---|---|---|
| Default safety-review data retention | 30 days, applied to top-tier models such as MythosCITE:E3 | 30 days, standard API defaultCITE:E4 |
| Latest quarterly revenue | $11.6 billion, nearly double the prior quarterCITE:E11 | $6.7 billion, up 18% quarter over quarterCITE:E11 |
OpenAI has made Private Safety Processing available to a limited set of customers for early testing, with broader rollout planned once the system maturesCITE:E9.
What this means: OpenAI and Anthropic currently run the same 30-day default retention window for safety review, but OpenAI's August 19 preview of Private Safety Processing arrived the same week the Wall Street Journal reported Anthropic's revenue reaching $11.6 billion against OpenAI's $6.7 billionCITE:E11. Commentary already ties the announcement to that competitive backdrop, describing it as a way to set OpenAI apart from Anthropic's retention policy and to give enterprises a stronger incentive to choose its platformCITE:E2CITE:E10. For now, the feature remains in early customer testing rather than general availabilityCITE:E9.
Private Safety Processing is a structural extension of OpenAI's existing zero data retention design, not a replacement for it: the current ZDR path requires healthcare and financial customers to apply and get approved before data collection stops, while the new system adds automated cross-interaction pattern detection and a customer-controlled encryption key storage option that leaves OpenAI without a key backup. The timing is notable — the preview landed the same week the Wall Street Journal reported Anthropic's quarterly revenue reaching $11.6 billion against OpenAI's $6.7 billion — and TechCrunch and Axios both read the feature as an enterprise incentive play rather than a purely technical safeguard. The unresolved question is adoption: OpenAI has confirmed only early-customer testing, with rollout timing tied to system maturity rather than a fixed date. The metric worth tracking is whether Private Safety Processing moves from early testing to general availability, and whether OpenAI publishes further detail on the client-controlled key storage option.
CK Hutchison Holdings (長江和記實業) filed international arbitration against Panama on August 20, 2026, seeking over US$1.5 billion after Panama's Supreme Court voided its port concession in January. Panama Ports Company (巴拿馬港口公司, PPC) separately seeks at least US$2 billion, while China and Hong Kong have pledged to defend the firms' rights and the US has welcomed the court ruling.
SK Hynix (SK海力士) announced on August 19 a ₩40 trillion share buyback-and-cancellation program and raised its 2025–2027 shareholder return target from up to 50% to over 50% of cumulative free cash flow, a day after its U.S. ADR fell 9.2% in a single session.
OpenAI announced Private Safety Processing on August 19, 2026, a zero-retention safety layer that scans encrypted, customer-controlled sessions for abuse without human access to content, previewed to select customers ahead of a September rollout. The move directly contrasts Anthropic's mandatory 30-day retention policy for its strongest models, which Forrester says overrides existing zero-retention agreements and which Anthropic itself has called a source of commercial risk.