Claude Sonnet 4.6 tops BullshitBench v2

This is the kind of benchmark win developers should care about: a model that pushes back on false assumptions is far more useful than one that sounds smooth while being wrong.

• –BullshitBench v2 is measuring epistemic discipline, so Claude’s lead here is about saying “that premise is false” at the right time, not just generating polished prose
• –Anthropic positions Sonnet 4.6 as a hybrid reasoning model for coding, agents, and 1M-context workflows, and this result supports that reliability-first pitch
• –Community reaction around the benchmark is notable because the story is not raw speed or style, but whether models can avoid confidently compounding user mistakes
• –For agent builders, better rejection of bad premises can mean fewer wasted tool calls, fewer hallucinated follow-on steps, and safer automation in production
• –Benchmark wins still need local evals, but this is a meaningful signal for teams optimizing around trust, not just benchmark theater