Claude was the only large language model cleared for classified Pentagon systems when Anthropic walked away from a $200 million contract on February 27. Six weeks later, OpenAI holds that slot under an "any lawful use" agreement with no published restrictions. If you're keeping score on AI safety outcomes, Anthropic's red lines produced a strictly worse result than the deal they refused to sign.
I get the instinct. Autonomous kill chains and mass domestic surveillance are legitimately terrifying capabilities to hand over. Anthropic's 2 specific objections, AI-selected drone strikes without human oversight and bulk geolocation tracking of U.S. citizens, are the kind of things that should make any engineer's stomach turn. Fair point: someone needed to say those words out loud in a negotiation room. But saying them and then losing the contract is not the same as preventing those capabilities from being deployed.
You Don't Get Credit for the Fork You Didn't Ship
Builders know this pattern. You refuse to merge a feature because it violates your architecture principles. The PM routes around you, hires a contractor, and the feature ships anyway, except now it has no tests, no error handling, and nobody who understands the codebase is maintaining it. That's what happened here.
Pentagon spokesperson Sean Parnell said in February that the military has "no interest" in mass surveillance or autonomous weapons without human involvement. Maybe that's true today. The problem is that OpenAI's contract doesn't encode those limits in writing. Anthropic's would have. The company that actually built safety constraints into its acceptable use policy is now blacklisted under the same statute used against Huawei and ZTE, while the company that said "sure, whatever's lawful" got the deal.
The D.C. Circuit's April 8 ruling made the math brutally clear: during an active military conflict, a court will not second-guess the Pentagon's choice of AI vendor. Anthropic can win on the legal merits eventually. By then the integration work will be done, the classified pipelines will run on OpenAI's models, and switching costs will make the original contract look like a rounding error.
Defensive Value Without Offensive Surrender
The irony is that Anthropic keeps proving its models are exactly what the government needs. Treasury Secretary Scott Bessent praised the company's Mythos model in early April for its cybersecurity capabilities against Chinese threats. Project Glasswing, Anthropic's controlled release of Mythos for defensive cyber, is real engineering with real constraints baked in. That's the kind of work that demonstrates you can serve national security without blanket capitulation.
So why didn't Anthropic negotiate from that position of strength instead of drawing absolute lines on 2 use cases the Pentagon publicly disavows? A contract that said "here are the 14 things Claude can do in classified settings, here's the audit trail, here's the kill switch" would have been harder to reject than a flat "no" on capabilities the Pentagon claims it doesn't want anyway. The 37 researchers from OpenAI and Google DeepMind who filed an amicus brief questioning legal safeguards clearly see the gap: the law doesn't prevent mission creep, and neither does walking away from the table.
Bessent estimates the U.S. AI lead over China at 3 to 6 months. That window is not wide enough for the country's most safety-focused AI lab to be locked out of defense work on principle while a less restrictive competitor fills the vacuum. The threat model isn't hypothetical. The military confirmed advanced AI use in the Iran conflict. These systems are in production now.
Anthropic built something genuinely good. Claude's safety research, its constitutional AI approach, its willingness to publish evaluation results. All of that matters more inside the Pentagon than outside it. The company that cares most about getting AI deployment right just ensured it has zero influence over how the Pentagon deploys AI. That's not a principled stand. It's a production outage with no rollback plan, and the only people who benefit are the ones who never wanted guardrails in the first place.
