OpenAI will offer zero data retention to eligible API customers from September, and has launched a monitoring service that looks for misuse without storing what customers send it, according to Computerworld.
The same update carried a quieter admission. The company has temporarily slowed its pace of scaling and paused reinforcement learning training for two weeks while it hardens its own research environment. Read together, the two announcements look less like a product release and more like a company tightening its controls before a regulator asks it to.
What OpenAI actually announced
The retention change is the concrete part. From September, API customers who qualify will be able to send data to OpenAI without it being stored, and the company says it will publish a technical white paper setting out how that works.
The rest is internal housekeeping made public. OpenAI paused reinforcement learning training for two weeks and put its largest planned frontier training run on hold, running smaller evaluations instead. It also says it red teamed its research environment, expanded monitoring, and added workload isolation, network isolation and continuous security testing.
How Private Safety Processing works
The monitoring service is called Private Safety Processing. Automated agents assess inputs and outputs across several conversations rather than checking each session on its own, which is meant to catch misuse that has been deliberately spread out over time to slip past a single session check.
If something trips the system, it emits a narrowly defined signal. OpenAI can then decide whether enforcement is warranted and may contact the customer for context, and the customer chooses whether to hand over any data. No human reads the conversations.
OpenAI drew the contrast with Anthropic itself. Anthropic keeps user sessions for 30 days for certain covered models, including its Mythos class models, and allows human review through a controlled access path that records who looked at what, as TechCrunch reported. Neither approach is obviously safer. They trade different things away.
The parts that are not settled
OpenAI has not said who qualifies. The eligibility criteria for zero data retention were left out of the announcement and are promised in the white paper, which means no one can yet check whether their own workload is covered.
There are also carve outs. Content flagged as child sexual abuse material is still retained for legal reporting, so zero retention is a policy with named exceptions rather than an absolute. The monitoring is not free either: OpenAI put the overhead at roughly 20 per cent of the inference compute being watched, varying by workload.
The commercial mechanics matter too. Some enterprises may need to supply their own API keys to obtain the contractual protections, which changes how access and margin flow through partners such as Microsoft and AWS. Analysts read the timing as positioning ahead of a possible public listing, and as self regulation offered before legislation arrives.

Why Australian businesses should care
Australian organisations that handle personal information have to be able to say where it goes and how long it is kept. Retention is usually the weakest clause in an AI vendor contract, because it is the one nobody reads until an incident forces them to.
It is worth separating two things that get confused. A promise to keep nothing is not a promise about where the processing happened, and the two carry different obligations. We covered that distinction in our guide to data residency choices.
For document heavy operations the stakes are higher. Contracts, invoices and client files carry far more identifiable information per request than a chat prompt does, which is why document automation work tends to run into retention questions first.
There is also a size question buried in the word eligible. Frontier model customers on direct contracts are the obvious candidates, which would leave most mid market Australian buyers depending on whatever their reseller has negotiated. Until the criteria are published, that is guesswork.
What to ask before moving sensitive work
Four questions cover most of it. Ask for the white paper and the eligibility criteria in writing. Ask for the full list of exceptions. Price the monitoring overhead into total cost rather than the token rate alone. And check whether your contract runs through OpenAI directly or through a reseller, because that appears to decide whether the protections reach you at all.
Defaults are the other thing to watch. Google spent last week explaining that Gemini reads Workspace data until an administrator switches it off, and the lesson generalises. What a vendor offers and what your account is actually set to are rarely the same thing on day one.
None of this is settled until the white paper lands. For Australian firms weighing whether to move sensitive work onto a frontier model API, the honest position today is that OpenAI has made a commitment worth taking seriously and not yet worth signing against.
