OpenAI Details Zero Data Retention for Frontier Models

Enterprise teams building on frontier models just got clarity on the question legal teams ask first: does anyone at the vendor see my prompts? OpenAI’s August 19 post, Offering Zero Data Retention for frontier models, reaffirms that qualifying API customers on frontier models keep true zero-retention guarantees — and previews a mechanism to catch abuse without breaking them.

Key facts

• ZDR, unchanged: qualifying Zero Data Retention customers get prompts and responses discarded once a request finishes processing; content is never used for training unless separately opted in.
• The one carve-out: apparent-CSAM detection and reporting legally requires retaining flagged images regardless of ZDR status — narrow and pre-existing.
• The preview: Private Safety Processing aims to detect cross-conversation abuse patterns through automated analysis while preserving ZDR.
• What’s next: a technical white paper is planned for September; the feature is not generally available.

Why this matters now

ZDR has been the difference between “we can build on this API” and “legal won’t sign off” for regulated industries — health, finance, and anything handling customer confidences. The tension: single-request retention policies can’t see multi-interaction risks like agents that ignore stop instructions or coordinated multi-account misuse. Vendors have historically resolved that by retaining more content for human review.

OpenAI’s answer keeps the opposite default: nothing retained, with automated pattern detection standing in for human review. A person enters the loop only if a customer chooses to escalate a flag. The contrast with Anthropic is structural — Claude’s enterprise terms retain covered-model content in a defined window so approved reviewers can inspect flagged sessions through a logged access path. Two defensible models of the same problem; buyers should know which one they’re signing.

The context

This lands mid-flight in OpenAI’s broader safety reset: the same week the company confirmed it paused parts of frontier RL training and put its largest planned run on hold over Astra cyber-capability concerns (our coverage). Anthropic publishes its own accounting in risk reports. Together they mark the shift from trust-me safety claims to auditable, contract-grade commitments. The new Agents API extends this infrastructure approach — managed sandboxes with ZDR-compatible partner environments.

What to watch

Whether the September white paper specifies exactly what the automated classifier sees and how false positives are contested — the compliance value of ZDR-plus-processing depends entirely on those details. And whether Anthropic responds with an equivalent no-retention tier for its own covered models.

Official source

OpenAI — Offering Zero Data Retention for frontier models (August 19, 2026).