OpenAI Keeps Zero Data Retention for Frontier Models

OpenAI is keeping Zero Data Retention (ZDR) alive for its frontier models, previewing a system that spots cross-session misuse without handing customer content to OpenAI staff. Announced August 19, Private Safety Processing is a direct counter to Anthropic's 30-day retention requirement on Claude Fable 5 and Mythos 5 — a split that matters most to security- and compliance-bound enterprise buyers.

What happened

On August 19, OpenAI previewed Private Safety Processing, a monitoring system that extends Zero Data Retention to frontier models. Under ZDR, OpenAI does not retain customer prompts or model responses after a request is processed, and customer content stays unavailable to OpenAI personnel for review (OpenAI, 2026).

Existing ZDR safety systems judge each interaction in isolation. As models take on longer multi-step work, OpenAI argues the most serious risks only become visible across a sequence of interactions. Private Safety Processing analyzes those patterns while keeping the underlying content on customer-controlled infrastructure. On detection, OpenAI receives only a limited signal naming the type of activity — never the prompts or responses that triggered it.

Rollout starts in September alongside a technical white paper, and applies to eligible API and enterprise customers, not consumer ChatGPT plans. Early customers reportedly include Microsoft and Databricks (The Next Web, 2026).

Why it matters

Anthropic now requires 30-day retention on its "Covered Models" — Fable 5 and Mythos 5 — a policy it said "will be unpopular" with customers who expected zero retention. OpenAI is betting it can deliver equivalent cross-session safety without the retention, which is a genuine fork for enterprises, governments, and regulated buyers whose compliance rules forbid handing prompts to a vendor.

Our directory verdicts on Claude Fable 5 and Mythos 5 both sit at conditional, partly on access and safety-classifier constraints. Anthropic's retention policy adds a data-governance term to that calculus: routing traffic to either model voids a prior zero-retention agreement for that workload.

What changes for you

If you weigh frontier models against a ZDR or data-residency mandate, this is the concrete differentiator to test in procurement. But hold the claim as a preview — OpenAI has not yet shown that signal-only monitoring catches what retention-based monitoring catches. Wait for the September white paper and early-customer results before treating it as proven.

FAQ

Does this apply to ChatGPT consumer users? No. The preview targets eligible API and enterprise customers; consumer ChatGPT plans are not included, and it adds no new privacy setting for regular ChatGPT users.

What's the actual difference between OpenAI and Anthropic? Anthropic retains Fable 5 and Mythos 5 prompts and outputs for 30 days — even for prior zero-retention customers — deleting them "in almost all cases" (Anthropic, 2026). OpenAI's system keeps content on customer-controlled infrastructure and sends OpenAI only a limited activity signal, not the underlying content.

Has OpenAI proven the signal-only approach works? Not yet. Private Safety Processing is a preview; the September white paper and independent testing will determine whether cross-session monitoring without retention matches what retention-based monitoring catches.