If OpenAI keeps nothing, how could it possibly detect someone misusing the API over many requests?

That is exactly the problem Private Safety Processing is meant to solve. OpenAI says the system identifies patterns across related interactions and surfaces only a limited signal describing the category and severity of the activity, without exposing the underlying prompts or responses to any human reviewer. OpenAI has not published the technical detail of how this works, and the system is not live — rollout and a white paper are stated for September 2026. Until then, the claim cannot be independently evaluated.

Does Anthropic retain enterprise data, and for how long?

For its Covered Models — currently Claude Fable 5 and Claude Mythos 5 — Anthropic states that for commercial zero-retention customers as of 9 June 2026 it requires limited data retention and review as part of its safety work, with prompts and outputs logged for 30 days on every platform where those models are offered. Anthropic's standard zero-retention arrangements outside that carve-out do not store inputs or outputs, though safety classifier results are still retained to enforce the usage policy. Bloomberg reported on 20 August 2026 that Anthropic plans to change these rules following enterprise concerns; no revised policy had been published at the time of writing.

Should this change which vendor I pick?

Not yet, on this basis alone. OpenAI's mechanism has not shipped and has not been described in evaluable detail, and Anthropic's policy is reportedly under revision. Both positions may look different by October. If you are signing now, the more durable protection is contractual — specifically, whether the vendor can unilaterally change retention terms mid-agreement, and whether the guarantee follows you onto the next model tier.

OpenAI Will Promise Not to Keep Your Data. Anthropic Says It Has To.

On 19 August OpenAI committed to Zero Data Retention for eligible API customers on frontier models, paired with Private Safety Processing — a system it says spots misuse across related interactions without exposing content to humans. Anthropic requires 30 days of retention on its Covered Models for exactly the reason OpenAI says it no longer needs any. Two labs, opposite public positions on whether safety monitoring requires seeing the data, and enterprise buyers now choosing between them on that axis.

ByDrafterDaily Editorial

PublishedAugust 23, 2026

Read7 min

AIEnterpriseTechnology

Questions

Frequently asked

4 answers

In practice it means prompts and outputs are not persisted after the request is processed, rather than never touching a server — the data must exist transiently in memory for the model to respond to it. It also rarely means nothing at all is kept. Both OpenAI and Anthropic retain derived artefacts such as safety classifier results or category-level safety signals even under zero-retention arrangements. If your compliance requirement covers derived metadata and not just content, read the specific terms rather than the label.

Go further

Track how AI governance becomes AI market structure

DrafterDaily covers the enterprise AI decisions that get made on policy rather than benchmarks — retention, residency, licensing and procurement. Analysis, not announcements.

More from AI on DrafterDaily

View all

ChatGPT Ads Can Target by Phone Number and Prompt: How List Matching Works, and What It Does Not Tell You

404 Media reports that advertisers can target ChatGPT users by email, phone number or device identifier, and by prompts. OpenAI says advertisers see only aggregates. Both can be true, and the gap between them is the story.

Oct 18, 20267 min read

Eight State Lawsuits Against OpenAI: Why Attorneys General Chose Consumer-Protection Law Over Product Liability

Attorneys general from eight states sued OpenAI on 14 October 2026. The legal route they picked, deceptive-practices statutes, matters more than the headline: it turns on what OpenAI said about safety, not on proving a defect.

Oct 17, 20268 min read

OpenAI Published 722 Math Manuscripts, Then Withdrew Three in a Day. What That Says About Checking AI Proofs

One sign error took out three of 722 manuscripts because later papers were built on earlier ones. The episode is a lesson in verification economics, and the verified share is still disputed.

Oct 15, 20267 min read

OpenAI Will Promise Not to Keep Your Data. Anthropic Says It Has To.