Sonnet 4.6

Anthropic just dropped Sonnet 4.6—their mid-tier AI model—and immediately got caught in a safety testing controversy. Here's what happened: they did a lighter assessment on Sonnet than they did on the flagship Opus model. Not a small difference. They skipped deeper extreme failure mode testing entirely. Their justification? The assumption that lower capability automatically means lower risk.

That's a bold assumption.

The system card—the document that transparency nerds actually read—disclosed this asymmetric evaluation depth. They openly admitted they didn't run Sonnet through the same rigorous safety hoops as Opus. The logic: "it's a less capable model, so what's the worst that could happen?" But here's the problem with that thinking—mid-tier models are the ones companies actually deploy at scale. They're the ones getting integrated into real products, real workflows, real user hands. If you're shipping something to millions of people, you might want to know exactly where it breaks.

Anthropic's Opus model got the full treatment—extreme failure mode testing, red-teaming, the works. Sonnet got a skip. The reasoning might technically check out on paper, but it raises questions about who's making the call on "capability" and whether that's actually a reliable proxy for risk. A model doesn't need to be superintelligent to cause real problems. It just needs to be confidently wrong in the right context.

Anthropic is based in San Francisco and continues to lead the AI safety conversation—though this particular chapter might have them backpedaling on the next system card.