Major Anthropic customers are voting with their feet. Palantir and Booz Allen have both started shifting workloads away from Anthropic's newest Fable models and toward OpenAI's Astra, all because of one thing: data retention policies.
Here's the issue. In June, Anthropic announced it would store customer data for 30 days when using Fable models, citing security reasons. Before that, the company had negotiated zero data retention deals with individual customers on a case-by-case basis. Now those customers are waiting. Some aren't willing to wait.
Anthropic says it's heard the complaints. Starting this fall, the company plans to launch Enterprise Frontier Safeguards, a program designed to offer zero data retention to "eligible" customers. The company has already identified some winners: Salesforce, Comcast, and Visa. They're getting zero data retention guarantees for Fable while the rest of the enterprise world watches and waits for the program to actually go live.
But here's the problem with making some customers wait: in AI time, weeks might as well be years. Palantir started offering OpenAI's Astra to its customers earlier this month with zero data retention locked in. That means enterprises using Palantir's platform get early access to Astra's capabilities weeks before they'll even be able to test Fable under acceptable terms.
One cybersecurity software company that works with the US government and major financial services firms made the switch explicit: it moved from Anthropic to OpenAI primarily because it could get zero data retention guarantees from OpenAI that Anthropic wouldn't provide. That's the kind of defection that matters in enterprise AI.
This is a remarkable moment. Anthropic has been dominating the enterprise space. Both Anthropic and OpenAI are preparing to go public. And yet OpenAI has managed to create a genuine competitive opening by simply offering what customers asked for without the waiting period.
The irony here is almost too perfect: Palantir CEO Alexander Karp is blasting Anthropic and OpenAI for using customer data to stay competitive... while Palantir built its entire business model on surveillance and data collection. The company that literally sells governments and corporations tools to harvest and weaponize personal data is suddenly very concerned about data privacy. It's like a tobacco exec lecturing about respiratory health.
Of course, both Anthropic and OpenAI have been under pressure to explain how their models learn from customer data. OpenAI faced particular scrutiny over whether its models trained on researchers' work to achieve a recent mathematical breakthrough (the company later denied it). The question of data use and model development is real and worth asking. But watching Palantir lead the charge feels less like principled advocacy and more like opportunity.
The broader lesson here is straightforward: customers care about data retention, and they'll switch if you make them wait. Anthropic's Enterprise Frontier Safeguards program is coming, but "coming" doesn't pay the bills if your competitor is shipping it now. OpenAI just turned a legacy weakness (enterprise trust) into a competitive advantage by being faster to deliver what enterprise customers actually want.
For Anthropic, the question isn't whether the program will eventually work. It's whether enough customers will still be there when it launches.