# OpenAI is keeping ZDR. The safety system is moving.

Sent August 20, 2026

On August 19, OpenAI restated Zero Data Retention for eligible API customers: prompts and responses are not retained after a request, staff do not review that content, and enterprise data is not used for training unless the customer opts in. Frontier models now run longer, multi-turn agentic work, where some risks only emerge across a sequence. OpenAI is previewing Private Safety Processing so ZDR can survive that requirement. For ZDR deployments, content remains on infrastructure the customer controls. OpenAI is also developing customer-controlled-key storage on its infrastructure, with automated systems returning narrow safety signals rather than the underlying prompts or responses.

## What it means

- [ZDR remains a procurement option](https://openai.com/index/offering-zero-data-retention-for-frontier-models/): Ask whether you are eligible, where content is stored, who holds the keys, and which signals leave your boundary.
- [Multi-turn safety is now a product](https://openai.com/index/offering-zero-data-retention-for-frontier-models/): If an agent keeps acting after a stop, that is the class of risk Private Safety Processing is designed to catch.
- [The CSAM exception remains](https://openai.com/index/offering-zero-data-retention-for-frontier-models/): Images flagged as possible CSAM are still retained for manual review and reporting, including in ZDR deployments.
- [A preview is not a signed control](https://openai.com/index/offering-zero-data-retention-for-frontier-models/): Private Safety Processing is in early-customer tests. OpenAI plans the technical white paper and rollout for September, while customer-controlled-key storage remains in development.

## The limit

We have not run this system. The public note does not specify eligibility, regional availability, or the exact signal schema. If you already use ZDR, wait for the September details before rewriting the DPA. This kind of policy change also will not appear as a new model ID.

## Analysis worth opening

- [Fast on a wafer is not a shortlist](https://benchlm.ai/blog/posts/cerebras-won-speed-not-the-shortlist): Cerebras is quicker than Groq on the models both host. The decision is which weights sit on the wafer, and whether an agent is actually waiting on tokens.
- [Groq is not a model](https://benchlm.ai/blog/posts/groq-is-not-a-model): After Nvidia licensed the chip and hired the team, what remains is a fast, narrow catalog. Most people searching “groq” still think they are buying a brain.
- [K3 is open. You still cannot run it.](https://benchlm.ai/blog/posts/kimi-k3-is-open-you-cannot-run-it): Moonshot’s 2.8T flagship sits next to the closed frontier. Open means downloadable shards and a license, not a machine you own.
- [Opus 5 ate the default. Fable is waiting.](https://benchlm.ai/blog/posts/opus-5-vs-fable-5): Claude Opus 5 is Fable-adjacent at $5/$25. That is the default now. Fable 5 still has a job.

## New on BenchLM

- [Track published retirement dates before the run fails](https://benchlm.ai/radar): Radar Pro keeps confirmed releases, published price changes, and deprecations together, with the original source attached. Current pricing and trial eligibility appear before checkout.

Archive copy reflects the rankings, prices, and availability stated when this issue was sent. Current pages may show newer evidence.

Canonical page: https://benchlm.ai/newsletter/issues/2026-08-20