This has been building for a while. The gap between 'AI safety' as a company value and real-world government contracts was always going to create friction. Still, it's hard to believe the US government's own AI capabilities are apparently behind what's publicly available. Although I know Opus 4.6 isn't the best model Anthropic is sitting on.