The two largest names in frontier AI have arrived at opposite answers to the same awkward question, and the disagreement is now a sales pitch. The question is this: how do you catch someone misusing a powerful model when the most dangerous behaviour is invisible in any single request? Both OpenAI and Anthropic agree the threat is real. A bad actor rarely announces intent in one prompt. They probe safeguards repeatedly, spread activity across accounts, and dress a genuine threat up as routine research. Spotting that pattern means looking across many interactions at once. The clash is over what a safety monitor is allowed to see while it looks.
OpenAI set out its answer on Wednesday in a post pointedly titled "Offering Zero Data Retention for frontier models." Its new system, Private Safety Processing, is designed to correlate activity across related interactions without any OpenAI employee ever reading the underlying prompts or responses. Automated systems do the analysis and emit only what the company calls a "narrowly defined signal" naming the type of activity involved, not the content itself. Customer data either stays on infrastructure the customer controls or sits on OpenAI servers encrypted with keys only the customer holds. OpenAI says its staff cannot reach the content in either case. A broader rollout and a technical white paper are promised for September, and consumer ChatGPT plans are unaffected.
Anthropic walked the other way. It recently told business users of its most capable models, including Fable 5 and Mythos 5, that it will retain their data for 30 days to support safety monitoring. The company conceded the move would be unpopular with customers used to zero retention, but argued that catching sophisticated, multi-step attacks requires holding the material long enough to study it. That is a coherent position: you cannot analyse in depth what you have already thrown away. OpenAI's counter is essentially an engineering claim, that clever cryptography and automated triage let you have the pattern-detection without the retention. Whether that holds up under a real adversary is what the September white paper will need to prove.
This is not a purely philosophical debate, because enterprise buyers are voting with their budgets. New figures from the corporate card company Ramp, covering more than 70,000 US businesses, show Anthropic still leading among paying business users at nearly 44 percent to OpenAI's roughly 40 percent as of July. But OpenAI, which lost that lead back in May, is now growing faster in the segment. Ramp's economist attributed part of the swing to OpenAI's newest model and, tellingly, said Anthropic's Fable tier "disappointed both in adoption and real-world application given price plus data retention requirements imposed by regulators." Data handling, once a footnote in procurement, has become a line item that moves market share.
The deeper point is that safety and privacy have stopped being a shared industry value and become a product decision, made differently by each lab and priced accordingly. For an enterprise weighing where to send its trade secrets, legal files and proprietary research, that is oddly clarifying. You no longer have to trust a vague promise to "keep data safe." You can read two concrete policies, one that forgets and one that remembers, and pick the trade-off you can live with. The labs have turned an unsolved safety problem into a menu, and for now the customer chooses.