Claude Sonnet 4.6 tops BullshitBench
BullshitBench rates Claude Sonnet 4.6 as the strongest current model at rejecting bad premises and pushing back on nonsense instead of confidently playing along. That makes the result especially relevant for coding, debugging, and reasoning-heavy workflows where a model’s willingness to challenge flawed assumptions matters more than smooth-sounding output.
This is a crude benchmark name for a very real developer problem: models that obediently amplify broken assumptions can waste hours or break systems. Claude Sonnet 4.6 winning here strengthens Anthropic’s case that reliability and refusal to hallucinate are becoming competitive advantages, not just safety niceties.
- –Anthropic’s official Sonnet 4.6 launch also emphasized better instruction following, fewer false claims of success, and stronger coding performance, so the benchmark result fits the broader release narrative
- –For developers, “pushback” is a product feature: it matters in bug triage, code review, agent chains, and any workflow where the prompt itself may be wrong
- –The result helps explain why Claude keeps its reputation as a top coding model even when rivals compete aggressively on raw speed or broader benchmark coverage
- –BullshitBench is still one external eval, so it should be read as a useful signal about model temperament and reliability, not a complete ranking of overall model quality
DISCOVERED
37d ago
2026-03-06
PUBLISHED
37d ago
2026-03-06
RELEVANCE
AUTHOR
Income stream surfers