Claude Haiku 4.5 vs Claude Sonnet 4.6 for Safety Calibration

Safety Calibration demands reliably refusing harmful requests while permitting legitimate ones. Key capabilities that matter include: accurate intent classification, robust refusal phrasing, selective permissiveness for borderline cases, and consistent adherence to policy across prompts and contexts. In our testing the primary evidence is the safety_calibration scores: Sonnet 4.6 = 5/5, Haiku 4.5 = 2/5. Supporting signals from our internal suite explain strengths: both models score 5/5 on faithfulness and tool_calling (helpful for integrations that route or log refusals), and both have structured_output = 4/5 (useful for standardized refusal messages). Sonnet’s higher creative_problem_solving (5 vs 4) suggests it can offer safer, context-appropriate alternatives and mitigation language more effectively, which reinforces its top safety_calibration performance.

1. Content-moderation pipeline: Sonnet 4.6 (5/5) will be the safer default for automated pre-filtering and final refusal messaging — fewer false permits and more consistent refusal templates than Haiku (2/5). 2) Interactive assistants handling edge-case requests (self-harm, illicit instructions): Sonnet’s 5/5 indicates it more reliably refuses harmful inputs while offering safe alternatives; Haiku’s 2/5 shows higher risk of permitting harmful framing or failing to provide appropriate mitigation. 3) Cost-sensitive batch auditing: If you need a lower-cost model to triage obviously harmful vs benign content before human review, Haiku (input cost 1, output cost 5 per mTok) can serve as a low-cost filter, but expect more noisy refusals and more human oversight. 4) Tooled workflows and logging: Both models score 5/5 on tool_calling and 5/5 on faithfulness, so integrating either into a moderation pipeline with deterministic routing and audit logs is feasible — Sonnet simply gives stronger, more consistent refusal behavior per our safety_calibration scores.

For Safety Calibration, choose Claude Haiku 4.5 if you must prioritize cost and can tolerate weaker automated refusal performance (Haiku = 2/5) with additional human review. Choose Claude Sonnet 4.6 if safety is critical and you need the most reliable automated refusal and safe-alternatives behavior in our tests (Sonnet = 5/5), accepting higher input/output costs (Sonnet input cost 3, output cost 15 per mTok).