Pentagon moves past Anthropic for a multi‑vendor AI backbone

The Pentagon has shifted away from Anthropic’s AI models, according to Palantir chief technology officer Shyam Sankar. In a high‑level interview, Sankar said the change has already happened and emphasizes it as part of a broader trend toward flexibility in choosing AI tools for military work. The Department reportedly moved to a mix of models from OpenAI, Google, and xAI, alongside open‑weight options from Nvidia. Developers inside the Pentagon can now pick among multiple AI providers to fit specific tasks, rather than relying on a single vendor.

Sankar framed this as a process rather than a single decision. He argued that military AI should be about selecting the right tool for the job, not about locking into one supplier. That’s aligned with Palantir’s business model, which centers on enabling customers to weave together different AI capabilities rather than pushing a single platform. The approach lets users adapt to changing capabilities or safety constraints without retooling their entire system.

The reported shift follows public and political scrutiny of Anthropic’s safety safeguards for military use. Anthropic has publicly engaged in debates about how its models should be constrained in sensitive applications. The Pentagon’s move away from Claude appears to reflect a desire for broader governance options and more direct control over which tools power critical tasks.

On governance and accountability, Sankar argued for clear rules and liability. He said existing laws should govern how AI is applied in defense contexts and suggested building a formal incident response and reporting regime to handle future issues. He also noted that Palantir did not join the White House AI accord, pointing to ongoing disagreements about which frameworks best suit government use of AI models.

The broader question is how the military will balance capability, safety, and transparency as it taps into a mix of commercial models. It’s not about a single “winner” but about a practical toolkit that can be updated as technologies evolve and as new safety standards emerge.

For readers, the practical takeaway is this: government agencies are increasingly comfortable stitching together AI from multiple sources to meet different needs. That means procurement and oversight may look very different in the next few years, as regulators and contractors work out who is responsible for what, and how quickly models can be swapped when safety or performance issues arise.