OpenAI is previewing a new system called Private Safety Processing to a small group of business customers. It is built to catch people misusing its AI models across many separate conversations, while keeping none of that customer's data in storage.
The timing is not an accident. Earlier this year, Anthropic told business customers that its most capable models would now hold onto their conversations for 30 days, so the company could review them for safety problems. Content flagged as risky can be kept for as long as two years. Anthropic calls these its covered models, and the list includes its Mythos and Fable model lines.
That change created real friction. Microsoft, which uses Anthropic's Claude models heavily across its own products, restricted its employees from using Claude Fable 5 while its legal team reviewed what the new retention rules meant for company data. When a company the size of Microsoft pulls back on a tool over a data policy, it tells you enterprise customers take this seriously, not as fine print but as a real business risk.
OpenAI's answer is to widen a system it already had. For years, OpenAI has offered what it calls Zero Data Retention, where an automated system checks each conversation for abuse without saving it afterward. Private Safety Processing stretches that same idea across multiple conversations at once. The reasoning is straightforward: someone trying to misuse AI for something harmful, like building malware, might spread their requests across several separate chats to avoid getting caught in any single one. The new system can connect those dots without a human ever reading the conversations. If it finds something concerning, it sends a narrow alert to OpenAI, which then decides whether to contact the customer directly.
Anthropic's approach is different but not reckless. Human reviewers can look at flagged conversations, but only through a restricted process, with every review logged in a way reviewers cannot edit or delete. Anthropic's argument is that some misuse is too subtle for a computer to catch alone.
Behind this privacy debate sits a much bigger financial rivalry. Anthropic told investors its annualized revenue run rate hit 65 billion dollars by the end of July, up sharply from 47 billion dollars in May. OpenAI's own second quarter revenue grew 18 percent to 6.7 billion dollars, a much slower pace than Anthropic's reported surge over the same stretch. Both companies are also preparing for a stock market listing, with OpenAI's private valuation near 852 billion dollars and some investors floating a 2 trillion dollar figure for Anthropic.
For any business signing a contract with either company, this is worth paying attention to. Data retention policy is no longer boilerplate on page thirty of a vendor agreement. It is becoming a genuine reason a company chooses one AI provider over another, especially for firms in insurance, law, healthcare, or finance where client data cannot be handled loosely. Expect both companies, and their competitors, to keep making privacy a selling point rather than a technical footnote.