A significant number of Anthropic's key enterprise clients are now restricting their usage of the latest Fable series of AI models. These customers had anticipated that the AI firm would not store the data generated from their model interactions, but the reality has fallen short of those expectations. This disconnect has prompted some clients to pivot their business toward OpenAI.

For instance, companies like Palantir and Booz Allen are expanding their deployment of OpenAI's Astra model, which comes with the benefit of a zero-data-retention option. Both firms have scaled back their use of the Fable models following Anthropic's June announcement that, for security reasons, it would store customer interaction data from Fable models for a period of 30 days. Prior to this policy adjustment in June, Anthropic had been able to offer zero-data-retention terms during individual negotiations with specific clients.

Later this fall, the AI company plans to launch a new initiative called Enterprise Frontier Safeguards (EFS), which would allow qualifying customers to actually experience zero data retention. It remains unclear what specific criteria companies must meet to gain eligibility for this program. However, Anthropic has already identified a group of customers to participate in building the program, including Salesforce, Comcast, and Visa. According to Anthropic's help documentation, some of these participants will receive temporary zero-data-retention guarantees for the Fable models before the program officially goes live.

Recently, Anthropic's customer data policy has become a major point of criticism across the AI laboratory sector. Palantir's CEO, Alexander Karp, has publicly condemned AI vendors, accusing them of using sensitive enterprise client data to compete against their own customers. This policy controversy has inadvertently created a benefit for OpenAI, which is aggressively expanding its enterprise business to strengthen its competitive position against Anthropic. Palantir opened access to the Astra model for its own customers earlier this month. This means a large number of enterprises using AI services through Palantir can now experience OpenAI's Astra capabilities weeks earlier, while they will still have to wait longer to test Anthropic's Fable model. In the AI industry, a lead of several weeks can be practically equivalent to years.

An executive from a cybersecurity software firm that serves the US government and large financial institutions noted that their company's end-customer applications have shifted from using Anthropic models to primarily relying on models like OpenAI's Astra. This shift is driven by the fact that OpenAI can provide guarantees of zero data retention, which Anthropic cannot currently offer. Of course, customers and the public alike are pressuring both Anthropic and OpenAI to disclose how their models utilize user data for training purposes. Previous reports claiming that OpenAI used third-party data to achieve a breakthrough in mathematics (which the company subsequently denied) have further amplified these external concerns. With the current momentum behind Claude Code products and both companies preparing for potential public listings in the near future, any competitive edge that OpenAI can secure is considered crucial.